Advances in Electrodermal Activity Processing With Applications For Mental Health

Alberto
Greco · Gaetano Valenza
Enzo Pasquale Scilingo
Advances in
Electrodermal
Activity Processing
with Applications
for Mental Health
From Heuristic Methods to Convex
Optimization
Advances in Electrodermal Activity Processing
with Applications for Mental Health
Alberto Greco • Gaetano Valenza
Advances in Electrodermal
Activity Processing
with Applications
for Mental Health
From Heuristic Methods to Convex
Optimization
123
Alberto Greco Gaetano Valenza
Department of Information Engineering Department of Information Engineering
Bioengineering and Robotics Bioengineering and Robotics
Research Center ‘E Piaggio’ Research Center ‘E Piaggio’
University of Pisa University of Pisa
Pisa, Italy Pisa, Italy

Department of Information Engineering
Bioengineering and Robotics
Research Center ‘E Piaggio’
University of Pisa
Pisa, Italy
ISBN 978-3-319-46704-7 ISBN 978-3-319-46705-4 (eBook)

DOI 10.1007/978-3-319-46705-4
Library of Congress Control Number: 2016952862
© Springer International Publishing AG 2016

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword
This book presents a critical review of methodological studies for the analysis of
the electrodermal activity (EDA), one of the most powerful noninvasive peripheral
measures of the autonomic nervous system (ANS) neural pathway. Through the
book, the author leads the reader from a thorough description of electrodermal
physiological phenomena to an advanced introduction and discussion of recom-
mended techniques for correct data collection and effective data analysis. Several
experimental setups are also presented.
Although the EDA signal is fairly easy to acquire and very informative, powerful
methodologies and efficient models are required to make meaningful inferences
on the dynamics at the central nervous system level. The book introduces and
emphasizes a novel computational model for EDA analysis that I have personally
promoted and coauthored. The method relies on rigorous mathematical techniques,
such as convex optimization, to provide an effective window on the ANS dynamics,
and it has been successfully applied in several experimental scenarios. EDA is a
source of many sensitive psychophysiological markers and it finds application in
several fields of research, such as psychology and medicine, as a viable indicator in
emotion assessment and pathological mood state recognition. Remarkably, the book
presents several experimental applications exploring different sensory channels
for emotion stimulation in both healthy subjects and bipolar patients with very
promising results.
I am confident the reader will find useful information on proper characterization
of EDA dynamics and how this can be applied to the rising fields of affective
computing and psychophysiology. The high technical content makes the book
attractive to anyone interested in signal processing, statistics, applied mathematics,
and physics.
The book is a valuable reference for active research scientists and postgraduate
students interested in methods at the interface of bioengineering and statistics. I
expect that this book will stimulate and encourage the use of such methods in
different fields of applied science.
Colchester, UK Dr. Luca Citi

August 2016
v
Preface
Electrodermal activity (EDA) can be considered one of the most common perceptual
channel, of the autonomic nervous system (ANS) dynamics and manifests itself as
changes in electrical properties of the skin. Several previous studies have shown
how EDA can be a very informative biomedical sign with high discriminant
power between different psychophysiological states, although in this case many
methodological issues arise. This book fervently shows how to retrieve much
reliable information from EDA, to investigate also the assessment of emotional
responses in healthy subjects and patients with pathological mood/mental states.
Throughout the chapters, in-depth methodological and applicative studies involving
EDA are described, including a critical review on the current state of the art.
Since continuous deconvolution analysis (CDA) has been recognized as one of the
mostly used methods for EDA analysis, we first show how to apply this model to
discern different affective states in healthy volunteers. Emotions were evoked using
multimodal standardized sets of pictures, sounds, caresses, and smells. Valence and
arousal levels of such emotions were identified as the principal dimensions of the
affective responses. The achieved results are consistent with the hypothesis that it is
possible to objectively study ANS dynamics involved in the emotional processing
by properly processing the EDA.
Furthermore, this book reports on a novel computational model for the EDA
analysis based on convex optimization methods. This model, hereinafter called
cvxEDA, describes the EDA as a sum of the phasic component, the tonic com-
ponent, and an additive white Gaussian noise term incorporating prediction errors,
as well as measurement errors and artifacts. CvxEDA is physiologically inspired
and overcomes the limitations of the heuristic solutions and post-processing steps
of the conventional approach. It is based on a rigorous methodology grounded on
Bayesian statistics, mathematical convex optimization, and sparsity. Building on our
previous CDA-based experimental results, outcomes of cvxEDA often demonstrate
higher accuracy than CDA while discerning elicited emotional states in healthy
subjects. When applied to EDA from psychiatric patients suffering from bipolar
vii
viii Preface
disorder, it is shown how EDA significantly changes according to different mood

states. This also allows using EDA phasic and tonic components as suitable markers
for discriminating pathological mood states in bipolar patients.
Pisa, Italy Alberto Greco

Gaetano Valenza
Acknowledgments
We would like to express our deepest and sincere gratitude to all the people who
contributed to data acquisition and analysis: Dr. Antonio Lanatá, Dr. Andrea Guidi,
Dr. Mimma Nardelli, Dr. Matteo Bianchi, Prof. Claudio Gentili, Dr. Nicola Vanello.
A special mention to Dr. Luca Citi for his fundamental contribution to our research
and for his foreword.
ix
Contents
1 Electrodermal Phenomena and Recording Techniques . . . . . . . . . . . . . . . . . . 1

1.1 Electrodermal Activity and Skin Conductance . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Anatomy of the Skin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Anatomy of Sweat Glands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Physiology of the Electrodermal System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4.1 Mechanisms of the Electrodermal
Electrophysiological Response. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.2 Genesis of the Electrodermal Response . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Recording Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5.1 Measurement Sites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Exemplary Electrodermal Activity Monitoring Devices . . . . . . . . . . . . . . 11
1.6.1 DC Source Front-End. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.6.2 AC Source Front-End. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6.3 Remote DC Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6.4 Wearable DC Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.6.5 Multi-Frequency Sensorized Glove . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2 Modeling for the Analysis of the EDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1 Mathematical Models of the EDA: An Overview . . . . . . . . . . . . . . . . . . . . . 19
2.2 EDA Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Conventional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Model-Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 CDA: Continuous Deconvolution Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.2 EDA Deconvolution Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.3 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 CvxEDA: A Convex Optimization Approach
to Electrodermal Activity Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.1 Convex Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.2 Model Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
xi
xii Contents
2.4.3 Observation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.4.4 Maximum a Posteriori Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5 Feature Extraction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.1 Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.2 Frequency Domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3 Evaluation of CDA and CvxEDA Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1 Synthetic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2 Experiment 1: Maximal Expiration Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3 Experiment 2: Visual Affective Stimuli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4 EDA Processing and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5 Experimental Evaluation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.5.1 Results on Synthetic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.5.2 Experiment 1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.5.3 Experiment 2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4 Emotions and Mood States: Modeling, Elicitation,
and Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1 Theory of Emotions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 Modeling Emotions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3 Autonomic Nervous System Correlates of Emotions . . . . . . . . . . . . . . . . . 48
4.4 Affective Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.5 Multi-Sensory Elicitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.6 Emotions and Mood Disorders: Bipolar Disorder . . . . . . . . . . . . . . . . . . . . . 53
5 Experimental Applications on Multi-Sensory Affective Stimulation. . . 55
5.1 Multi-Sensory Experimental Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.2 Classification Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2.1 Paired Within-Rank K-NN Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2.2 Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3 Affective Visual Elicitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3.1 Experimental Protocol of Affective Visual Elicitation . . . . . . . . 61
5.3.2 Classification of Visual Arousal and Valence Levels . . . . . . . . . 62
5.4 Affective Sound Elicitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.4.1 Subject Recruitment, Experimental Protocol
and Acquisition Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.4.2 Feature Extraction and Statistical Analysis. . . . . . . . . . . . . . . . . . . . 64
5.4.3 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.5 Affective Touch Elicitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.5.1 A Device for Caress-Like Haptic Stimuli . . . . . . . . . . . . . . . . . . . . . 70
and Acquisition Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.5.3 Feature Extraction, Performance Metrics,
and Statistical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.5.4 Statistical Results of the Self Assessment Questionnaire . . . . . 75
5.5.5 Experimental Results of Tactile Stimulation . . . . . . . . . . . . . . . . . . 76
Contents xiii
5.6 Affective Olfactory Elicitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

and Acquisition Set-Up of the Olfactory Stimulation . . . . . . . . . 88
5.6.2 Feature Extraction and Statistical Analysis. . . . . . . . . . . . . . . . . . . . 90
5.6.3 Statistical Analysis on Self-Assessment
Questionnaire Scores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.6.4 Statistical Analysis and Classification of
Olfactory Valence Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.6.5 Dataset Reduction and Gender Analysis. . . . . . . . . . . . . . . . . . . . . . . 92
5.7 Assessment of Mood States in Bipolar Patients Using EDA. . . . . . . . . . 93
5.7.1 Patient Recruitment and Experimental Protocol . . . . . . . . . . . . . . 94
5.7.2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.8 Changing Source Oscillations of Skin Admittance:
A Study in the Frequency Domain with Application
on Emotion Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.8.1 Experimental Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.8.2 EDA Analysis and Classification Procedure . . . . . . . . . . . . . . . . . . 106
5.8.3 Classification Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.1 Future Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Acronyms and Symbols
The following table shows the most used and important acronyms.
Statement Acronym
Autonomic nervous system ANS
Central nervous system CNS
Electrodermal activity EDA
Electrodermal response EDR
Electrodermal level EDL
Skin conductance SC
Skin conductance response SCR
Skin conductance level SCL
Non specific skin conductance response NsSCR
Heart rate variability HRV
International Affective Picture System IAPS
Circumplex model of affect CMA
International Affective Digital Sounds IADS
Continuous deconvolution analysis CDA
Signal-to-noise ratio SNR
Finite impulse response FIR
Infinite impulse response IIR
Impulse response function IRF
Auto-regressive moving average ARMA
Auto-regressive AR
Moving average MA
Area under the curve AUC
Respiratory volume RV
Respiration activity RSP
Skin temperature ST
Pupil diameter PD
xv
xvi Acronyms and Symbols
Impedance cardiogram ICG

Heart sound HS
Linear discriminant analysis LDA
Artificial neural network ANN
Support vector machine SVM
Canonical correlation analysis CCA
Stepwise discriminant analysis SDA
Classification and regression tree CART
Quadratic discriminant analysis QDC
Introduction
This book is intended to provide an exhaustive description of the electrodermal

activity (EDA), from a deep insight onto the physiological foundations to ad hoc
algorithmic methods to analyze it. Expected audience ranges from researchers
with expertise in signal processing who would like to approach EDA analysis for
their first time, to experienced EDA researchers aimed to take into account recent
advances in EDA sparse modeling. Proper links to MATLAB software for EDA
analysis are also provided (see Chap. 2). Our principal aim is to show how EDA
can be at the center of breakthrough investigations involving the autonomic nervous
system (ANS) activity, being also a source of reliable and effective biomarkers of
healthy affective responses and pathological mood/mental states.
EDA manifests itself as a change in electrical properties of the skin, i.e., skin
conductance (SC). There are two main components of EDA having different time
scales and relationships with exogenous stimuli: the tonic and phasic components.
In the first part of the book, we describe the electrodermal physiological
phenomena underlying SC variations. Moreover, we include a critical review on the
current state of the art concerning EDA application, analysis methods, and recording
systems for both laboratory settings and ecological scenarios. The description
of a recently proposed recording system which uses different frequencies for
the demodulation of EDA components is also emphasized. Importantly, despite
the widespread use of EDA device-related measurements, the actual biological
phenomena underlying EDA (i.e., skin sympathetic nerve activity) remain unknown.
Therefore, in the last decades, several mathematical models were developed to
overcome this limitation, trying to investigate on how ANS activity regulates
the EDA dynamics. In this book, we rely on the classical model describing SC
as the sum of three terms: the phasic component, the tonic component, and an
additive white Gaussian noise term incorporating model prediction errors as well
as measurement errors and artifacts.
In this part of the book, we also emphasize a recently proposed, physiologically
inspired EDA model based on a rigorous mathematical approach, grounded on
Bayesian statistics, convex optimization, and sparsity. The phasic component is
seen as the result of a convolution between a bioinspired bi-exponential impulse
xvii
xviii Introduction
response function (IRF) and a sparse signal representing the sudomotor nerve
activity, which is part of the ANS. The IRF is modeled as an IIR filter allowing a
much more compact and non-banded matrix representation increasing the accuracy
and reducing the computational cost. Unlike previous algorithms in the literature,
this model incorporates the intrinsic physiological characteristics of EDA without
necessarily resorting to heuristics and ad hoc solutions, thanks to the presence
and definition of prior probabilities for the phasic and tonic signals. Results were
compared to those obtained through the continuous deconvolution analysis (CDA)
model [1], a method that performs a deterministic inversion of the peripheral
model. The proposed method showed good performance confirming a promising
applicability in the field of affective computing as well as of mental health.
In the second part of the book, we report on several EDA application sce-
narios, especially related to two specific research fields: emotion recognition and
assessment of mood/mental disorder. Indeed, emotions and mental disorders are
strictly and intrinsically interrelated; therefore, when emotions are dysregulated,
mental health is not guaranteed. Affective experiences accompany all cognitive
processes and social activities even in the case of psychopathologies [2, 3].
Moreover, prevalent theories affirm that the emotional processes can have primacy
over cognition [4]. As an example, the regulation process of emotions is crucial
in the occurrence and control of major depressive episodes, and some theoretical
views of depression are based on emotion changes which have implications in the
assessment, treatment, and prevention of the pathology [5]. Another well-known
relationship between emotions and mental disorders regards anxiety [6] as well as
brain damages of emotional processing areas and decision-making process [7].
In this part of the book, several experimental results gathered from testing EDA
models to robustness to noise, ability to separate and identify exogenous stimuli,
and capability of properly describing the activity of the autonomic nervous system
in response to specific affective elicitation are reported in detail. Concerning the
affective elicitation paradigm, we show exemplary applications of EDA modeling
on data gathered from healthy subjects undergoing multimodal affective elicitation,
where visual, auditory, olfactory, and tactile stimuli were investigated. Concerning
the mental health scenario, EDA analysis was employed to assess patients with
bipolar disorder [8–10], who experienced depressive and manic or hypomanic
episodes. Data used for this study were acquired in the frame of a European
collaborative project called PSYCHE (personalized monitoring systems for care in
mental health) [8, 11].
Chapter 1
Electrodermal Phenomena and Recording
Techniques
1.1 Electrodermal Activity and Skin Conductance
Electrodermal activity (EDA) is the general term used to define autonomic changes
in the electrical properties of the skin. One of the most frequently used measures of
EDA is skin conductance (SC), which can be quantified by applying an electrical
potential between two points of skin contact, usually the medial or distal phalanxes
of the non-dominant hand, and measuring the resulting electric current between
them. Electrodermal signals are a manifestation of the activity in eccrine sweat
glands that are innervated by the sympathetic branch of the autonomic nervous
system (ANS), mainly by the sudomotor nerves [12]. Indeed, when the sudomotor
nerves stimulate the production of sweat, the conductivity measured on the skin
surface changes as a result of sweat secretion and variations in ionic permeability of
sweat gland membranes [13–15]. The EDA is comprised of two main components,
having different time scales and relationships with the triggering stimuli: tonic and
phasic. The tonic EDA is given by the skin conductance level (SCL) which repre-
sents the slow-varying baseline level of the SC. Variations in the SCL are thought
to reflect slow changes in the ANS dynamics. The phasic EDA is represented by
a fast changing component, called skin conductance response (SCR), reflecting the
evoked response of the eccrine sweat glands to an external stimulus. The SCR is
defined as the SC transient arising within a predefined window (1–5 s) after the
stimulus onset and satisfying a minimum amplitude criterion (0.05 S) [16]. Recent
evidences suggest that these two components rely on different neural mechanisms
[17] and, consequently, that both convey relevant and non-redundant information
about the ANS activity.
The SC can be easily measured by applying a constant 0.5 V potential across two
skin-contact points, usually on the surface of hands or more specifically of fingers,
where there is a high concentration of eccrine sweat glands [18, 19]. EDA is used in a
wide range of experimental setups because it is a relatively straightforward measure
© Springer International Publishing AG 2016 1

A. Greco et al., Advances in Electrodermal Activity Processing
with Applications for Mental Health, DOI 10.1007/978-3-319-46705-4_1
2 1 Electrodermal Phenomena and Recording Techniques
Table 1.1 Structure of the Cutis Epidermis Stratum corneum

layers of the skin
Stratum lucidum
Stratum granulosum
Stratum spinosum
Stratum germinativum
Dermis Stratum papillare
Stratum reticulare
Subcutis Hypodermis
providing valuable information about the ANS response to a broad range of external
stimuli. In particular, SC analysis is commonly used to quantify the levels of arousal
associated with emotional and cognitive processes [17, 20, 21].
Although sweating is primarily a means of thermoregulation, sweat glands
located on the palmar and plantar (glabrous) surfaces possibly evolved to increase
grip and enhance sensitivity, and may be more responsive to psychologically
significant stimuli than to thermal ones [13, 15]. This relationship between EDA,
ANS, and psychological stimuli—together with the relative ease of measurement—
makes this physiological signal widely popular in neuroscience research, including
information processing, quantification of arousal levels during emotional and cogni-
tive processes, and clinical research examining predictors and correlates of normal
and pathological behaviour [17, 20–22], such as psychopathology, personality
disorders, conditioning, and neuropsychology.
1.2 Anatomy of the Skin
The main role of the skin is to protect the body from environmental threats such
as temperature, chemical, mechanical and infectious agents. It acts as a selective
barrier, and provides sense functions thanks to the mechanoreceptors, nociceptors,
and thermoreceptors. As a selective barrier the skin acts an other important role: the
regulation of perspiration both to prevent the body from drying out and to control
the emission of fluid thanks to the sweat glands.
The skin is not a single organ but consists of a complex set of organs and it is
possible to distinguish several layers (see Table 1.1).
The cutis is comprised of two sub-layers, the dermis and the epidermis (see
Fig. 1.1). The epidermis is the outermost layer of the skin and consists of the
epithelial tissue, which becomes progressively hornier closer to the surface. At
its base there are numerous layers of cells that reproduce continuously and move
towards the outer layers in order to replace those that die and fall off. The
epidermis does not contain blood vessels, and it is nourished by diffusion through
the underlying dermis. In particular, the stratum corneum plays an important role in
the electrothermal phenomena: generally this layer is dry, but becomes wet in the
presence of sweat.
1.2 Anatomy of the Skin 3
Stratum corneum
Stratum lucidum
Stratum granulosum
Stratum spinosum
Stratum germinativum
Stratum
papillare
Dermis
Stratum
reticulare
Subcutis
(Hypodermis)
Fig. 1.1 Layered composition of the glabrous human skin. An eccrine sweat gland, encircled by
its glomerulus, together with its straight dermal and irregularly coiled (helical) epidermal duct
(labeled acrosyringium), is shown in cross section. A part of the reticular layer has been omitted
due to its size in relation to the rest of the skin. Courtesy of [23]
The epidermis is relatively thin in comparison with the deeper-lying dermis. This
is located under the layer of the epidermis and consists of a connective tissue, which
acts as a cushion for the mechanical trauma of the skin. The dermis is tightly
connected to the epidermis by a basement membrane and contains many nerve
endings (mechanoreceptors) that provide the sense of touch and heat. It also contains
blood vessels, lymph vessels, sebaceous glands, sweat glands, hair follicles and
apocryphal [24–27]. The hypodermis (subcutis) separates the cutis from the deeper-
lying tissue and is composed of loose connective tissue. It contains the secretory
part of the sweat glands, appearing as a glomerulus (Fig. 1.1), as well as fatty tissue,
and the larger vessels which supply the body surface [23].
There are regional differences not only in the skin’s vertical layering, but also
in its horizontal structure. In the early stages of embryonic development, different
patterns of skin are formed by either ridge formation or folding into polygonal
structures [24]. Thus, the different types of skin are referred to as ridged skin
and polygonal skin. Specifically, ridged skin is seen only on the palms and soles.
The sweat gland ducts usually enter the epidermis at the nadir of the ridges. The
ridged skin is glabrous (hairless) and has no sebaceous nor scent glands. The rest
of the body is covered by polygonal skin patterns. Unlike ridged skin, the ducts of
sweat and scent glands enter the epidermis at the higher parts of the skin. Hairs and
sebaceous glands, however, are located in the channels of the polygonal skin.
All these layers have different properties depending on body sites. Concerning
the EDA, palms and fingers of hands are very important for their specificity
for emotional sweating. They are characterized by a very thin epidermis, i.e.
approximately 1 mm while it is ordinarily 50–200 mm.
1.3 Anatomy of Sweat Glands
The sweat glands cover almost the entire surface of the human body; they amount
to more than three millions, increasing by a factor of 7 from birth to adulthood.
They are present in high concentrations in specific areas such as palms, forehead
and soles, and in low concentrations on arms, trunk, legs [28], whereas there are
areas in which they are totally missing, such as lips or the inner ear channel [29].
Sweat glands are considered as exocrine glands because they secrete directly onto
the skin’s surface. Concerning the characteristics of the sweat secretion, sweat
glands are classified into eccrine and apocrine. Apocrine sweat glands play only
a negligible role with respect to the total amount of sweating [30] and respect to
the electrodermal activity too, and they are distributed especially on the breast, the
axillary, circumanal and genital regions [24, 31]. The eccrine sweat glands, instead,
play the most important role in the electrodermal activity phenomena. They are
the majority of sweat glands in human beings and are present in almost all the
surfaces of the body. Their secretions, i.e. eccrine sweat, do not contain noticeable
amounts of cytoplasm from the glandular cells, and they are not continuous, but
depend on various stimuli that affect also the regions in which they take place.
Heat, for example, causes sweating mainly on forehead, neck, back, chest and
back of hands [24]. Instead, emotional stimuli affect armpit regions, sides of the
torso, palms of hands and soles of feet. In 24 h the human body can produce
a huge amount of eccrine sweat, which can be up to 10–12 l [32]. The gland
activity is controlled by the hypothalamus and, as stated above, the sweat glands
are innervated by the fibers belonging to the sympathetic nervous system. They
are highly vascularized and innervated by a dense network of nerves, which are
both cholinergic and adrenergic. The secretory part of eccrine glands is innervated
only by the sympathetic cholinergic nerves: that means that they use exclusively
acetylcholine as a synaptic transmitter, which is produced and used on site.
1.4 Physiology of the Electrodermal System 5
As a matter of fact, although sweating occurs in response to extremely different

stimuli, (i.e., thermoregulation, emotional response (sudomotor nerve activity), or
the perspiratio insensibilis, (necessary to body water balance), the different behavior
of sweat glands is not related to structural and anatomical differences of the gland
itself, but only to the region where it is located. The eccrine glands are classified into
three groups according to the different behaviors corresponding to the anatomical
regions where they belong (of note, this distinction is not, of course, absolute, as in
most of biological phenomena):
• The eccrine glands on palms and soles respond almost exclusively to emotional
stimuli, such as pain, fear and anxiety, mediated directly by the brain structures
involved in emotion; or they respond to particular physiological stimuli, such as
movement and deep breathing.
• The eccrine glands of face, arms and forearms occupy an intermediate position
between the thermo-regulatory and emotional sweating, responding moderately
to psychic and thermal stimuli.
• The eccrine glands of the other areas of the body are almost exclusively involved
in the temperature control function.
1.4 Physiology of the Electrodermal System
Although the local processes underlying the EDA in the skin are well known, the
central origins of EDA are still under study. EDA is the result of the interaction
between local processes in the skin and sympathetic nervous system activity. At least
two different CNS sources have been identified as areas controlling the sudomotor
activity that leads to the electrodermal changes (see Fig. 1.2).
However, they use the same peripheral sudomotor efferents to the sweat glands
as a common final pathway. Specific central sudomotor pathways are not well
established, and even with the aid of brain imaging, no final breakthrough has
been reached in clarifying the central elicitation or inhibition of electrodermal
phenomena [23].
The hypothalamus is the area of the brain that directly regulates the secretion
of sweat; it is the main center of thermoregulation, then it is supposed to play
an important role in the origin of the EDA signal. The hypothalamus functions
are influenced by different brain structures, which are part of the limbic system.
Moreover, also the cortical area interacts with the limbic system, which can also
affect the hypothalamus activity and, consequently, the EDA. As a matter of fact,
previous studies [33, 34] have shown a decrease or the disappearance of the skin
conductance response to an emotional stimulus in subjects with lateral-frontal
lesions, comparing with other cortical areas lesions. Therefore, the sudomotor reflex
seems to be under the complete control of the limbic system and brain structures
associated with it. Specifically, the amygdala and the hippocampus are the limbic
structures mostly involved in the control of the hypothalamic functions related to
Premotor
cortex (Area 6)
Ci
Limble System
Ci = Cingulate gyrus C Basal Ganglia
A = Anterior thalamus C = Caudate nucleus
Fo = Fornix P = Putamen
A Fo
M = Medial part of the
Hi = Hippocampus
Hy = Hypothalamus P }
L = Lateral pallidum
L
M
Hy
Hi
1 2
3
RF = Reticular RF
formation Medulla
Sympathetic
anterolateral
pathway
Spinal cord
Fig. 1.2 Central nervous system elicitation of EDA in humans. 1: ipsilateral influences from the
limbic system via hypothalamic thermoregulatory areas (EDA1); 2: contralateral influences from
premotor cortical and basal ganglia areas (EDA2); 3: reticular influences. Courtesy of [23]
the sudomotor nerve activity. The amygdala should play an excitatory function,
whereas the hippocampus plays an inhibitory function. Specifically, the amygdala
plays a key role in sweating control as response to emotional stimuli. This because
of the involvement in the behavioral pattern memory (emotions behavior, social
behavior, endocrine and autonomous functions). In 1965 Bagshaw [35], removing
the amygdala from the brain of some monkeys, found out that the activation of
the EDA does not terminate with the limbic-hypothalamic activity, but also occurs
as a result of phenomena such as deep breathing and movements. This suggested
the presence of at least two other sources: the premotor cortical and basal ganglia
1.4 Physiology of the Electrodermal System 7
and the reticular system. The first consists of premotor cortical areas, whose fibers
for the transmission of pulses are found in skeletal muscle connections close to
the fibers that control the secretion of sweat. In fact, when these cortical areas
are naturally or electrically stimulated, or removed, an intense sweating can be
observed. The reticular system is a nuclear and fiber network in the inner part of
the brain. It plays a very important role in various phenomena, such as keeping the
alertness, processing of sensory stimuli, adjustment of spinal reflexes. Therefore,
the reticular system can trigger and modulate the EDA.
There are lots of empirical evidences that electrodermal activities are gener-
ated by sweat glands in conjunction with epidermal membrane processes. Sweat
secretion leads to two phenomena: the filling of sweat ducts and the moistening of
the relatively dry upper epidermal layer, the stratum corneum. When the ducts are
filled, the skin conductance increases due to shunts through the epidermal barrier
that connect the surface of the skin with the highly conductive dermal tissue. On
the other side, the moistening of the corneum increases skin conductance thanks
to salty sweat. In the literature, there are several simple electrical models that
can describe those purely resistive properties of the electrodermal system. In 1983
Edelberg proposed a resistive model of EDA, which resembled the simplified form
of the Montagu-Coles model [36] as depicted in the right-hand part of Fig. 1.3. He
regarded the corneum and the sweat duct as resistors in parallel, connected in series
with a resistor which includes some corneal and all subcorneal structures, except
the sweat gland lumen. Edelberg [37] also did not include the frequently discussed
active epidermal membrane, because of its hitherto uncertain role in contributing
to conductance changes [38]. In case of exosomatic recording, such a membrane
would act as a capacitor lying in parallel to the resistors of the corneum and sweat
duct.
1.4.1 Mechanisms of the Electrodermal Electrophysiological

Response
By applying a potential difference between two sites of the skin, it is possible

to observe a flow of current through it. This is due to the movement of free
ions present in the skin structures. Blood, sweat ducts and interstitial fluid have
a different conductivity because of their different ionic concentration. The dermis
and the subcutaneous layer act as good conductors because of the bloodstream,
the lymphatic system, and the interstitial fluid. On the other hand the corneum
shows relative impermeability to water and acts as a barrier for the diffusion of
the ions. This barrier, however, is traversed by the ducts of sweat glands, which
allow the diffusion of both sweat and ions. The slow changes in the EDA due
to this mechanism have been postulated as the cause of the tonic component. In
addition to the resistive properties, the skin has capacitive properties. They derive
from the cell membranes, which may show similar characteristics to the capacitors.
Fig. 1.3 Left-hand panel: Electrical equivalent circuit for the skin, according to Montagu and
Coles [36]. R1: resistance of the dermis and the body core. R2: resistance of the stratum corneum.
r1, . . . , rn: connectable resistances of the sweat gland ducts. C: capacitive element. Right-hand
panel: Simplified Montagu-Coles model. R: variable resistance resulting from sweat gland ducts.
Left-hand panel from [36]. Courtesy of [23]
In fact, the membranes can store electric energy and become potential generators.
The membranes that behave as polarized capacitors are presumably mainly localized
in the secretory activity of the sweat glands, in the transition area between the dermis
and the epidermis and into the epidermis. In conclusion, the skin can be interpreted
as an electrical network formed by RC circuits in series and parallel. We can assume
that the capacitive elements can be loaded very quickly, but once that they are fully
charged, they can not scroll other current and therefore they can not further influence
the EDA.
1.4.2 Genesis of the Electrodermal Response
The sequence of events that occur during the electrodermal response in the
epidermal duct of a sweat gland is shown in Fig. 1.4.
• at the beginning of the response, if the surface layer of the stratum corneum is
well hydrated, the pore sudoriparo and the distal part of the ductus sudoriparo are
closed under the effect of pressure exerted by the surrounding stratum corneum
(approximately 20 mmHg). In fact, the conductance does not increase.
• If the sweat fills the duct, it occurs higher conductivity of the duct itself and thus
measuring an increase in conductance in the horny layer (composed of dead cells
now dried).
1.5 Recording Systems 9
Fig. 1.4 Sequence of events during the electrodermal response
• If the rate of secretion is high enough, it can generate a pressure intraductal

greater than that of the stratum corneum. In this case there will be a pore opening
of the duct, which is traversed by the sweat. When the end portion of the duct is
opened, the conductance increases again. The sweat will be pushed out through
the pore but will also continue to press laterally in the stratum corneum. The
increased hydration (laterally in the stratum corneum and axially through the
duct) is the most important factor that contributes to a rapid increase of the
conductance and is supposed to be responsible for “phasic component” signal
electrodermal.
• When the sweat leaves the duct intraductal pressure decreases, unless the rate of
secretion is such as to compensate for the lost volume. If this becomes smaller
than the outside, the pore and the terminal part of the duct close again, causing a
rapid decrease of the conductance (see an example of SC signal during affective
elicitation in Fig. 1.5.
1.5 Recording Systems
Three standard methodologies are usually employed to record the EDA. The first
one is called endosomatic measurement. This is rarely used and will not be detailed
in this book. It consists in measuring, directly on the skin, the potential difference
between two skin sites, in a passive way. It does not need special amplifiers and
coupling electrical circuits. Although it is a quite unknown bioprocess, it is accepted
that changes in skin potential during sympathetic activity may be provoked by the
sodium reabsorption across the duct walls and the consequential change of the ionic
potential in the sweat ducts [15].
2
1
0
6
5
μS
4
3
2
1
0
0 50 100 150 200 250 300 350 400 450 500
t [s]
Fig. 1.5 Representative SC signal during affective stimulation. (Of note, the first 150 s correspond
to a resting phase)
The other two methods are based on an exosomatic approach, i.e., a small exter-
nal current is directly injected into the skin. In this regard, two different methods are
used, specifically, the direct current method (DC) and the alternating current method
(AC) at different frequency levels. Generally, they perform a measure of resistance
(impedance) or conductance (admittance: “conductance C j susceptance”), where
j is the imaginary unit. When DC source is used resistance and conductance are
inverse, while when we use AC source the inverse paradigm is not valid, due to
changes in the equivalent electrical model of the skin. In fact, it changes from a
parallel resistor-capacitor circuit (where, in DC regime, resistance and conductance
are inverse) to a series resistor-capacitor circuit [39].
Although AC methods allow measuring capacity changes in the electrodermal
responses, DC procedures are the most implemented [40]. Within DC methodology,
even if much effort has been spent to standardize, an agreement concerning
the use of either constant-voltage or constant-current scheme has not yet been
achieved. Therefore, different designs can be found. In constant-voltage sources the
conductance of the skin can be directly measured as an output of the circuit without
the need of any further transformation. However, constant-current sources provide
more stability and exhibit less tolerance, but in this method much attention must be
payed to possible damage to the sweat ducts due to the injected current through a
small area of the skin [39].
Unfortunately, a very low number of studies have been published on the
difference among AC and DC stimulation in EDA measurement. One of the
most interesting study [40] showed that using an AC source at 88 Hz, the major
contribution to EDA is given by the conductance term and not by the susceptance.
More specifically, the authors found that, at the instant of the conductance response,
there were no susceptance responses, and this could indicate the absence of a not
significant capacitance in the sweat ducts.
1.6 Exemplary Electrodermal Activity Monitoring Devices 11
Moreover, to our knowledge, no systems able to perform both DC and AC

electrical stimulations can be found in the current literature. In this book a new
device enabling both modalities will be described.
The EDA can be a rich source of information if monitored continuously in
an ecological scenario. In this view, the use of a wearable monitoring system
is crucial. As a matter of fact, wearable sensors are greatly valued due to their
comfort, portability, non-invasiveness, and their wireless communication capabil-
ities with either a computer, a mobile embedded system or other wearable sensors
[39, 41–45].
Starting from the above considerations and from the recommendations provided
in [46–48], in this book we described in detail a textile wearable system (see
Sect. 1.6.5), which is able to perform an exosomatic EDA measurement using both
AC and DC methods. Moreover, in Sect. 5.8, this novel device is used to investigate
whether the admittance contribution at different frequency sources (in the range
from DC to 1 kHz) could affect the capability of EDA of inferring the central state
during emotional stimulation.
1.5.1 Measurement Sites
As mentioned above, conventionally, EDA is measured at palmar and finger sites.

Suggested locations for electrode placement are given in Fig. 1.6. In general, the
electrodes used are made of Ag/AgCl and they require the use of a suitable
conductive paste. The availability of more unobtrusive forms of EDA sensing has
increased with the proliferation of mobile computing [49]. There are several mobile
EDA sensors in wristband or other wearable devices.
van Dooren et al. [50] investigated the relative strength of EDA responses on
16 body locations. They used a conventional EDA device for all measurements
and aroused the subjects via emotional film segments. The study concluded that,
considering the finger phalanxes as gold standard, EDA measurements at the soles
of foot were most similar with those of the fingertips. In contrast, arm, back, and
armpit traces differed most from the finger trace. Poh et al. [51] compared the
measurements of mobile EDA sensors on wrists with measurements acquired using
conventional EDA sensors on fingers. An important result in this latter study is the
evidence that the distal forearm is a viable alternative to the traditional palmar site
for EDA measurements.
1.6 Exemplary Electrodermal Activity Monitoring Devices
So far, the hitherto nearly exclusive use of DC methods in exosomatic EDA

recording prevented the development of AC recording techniques therefore almost
the totality of remote and wearable devices has been based on the DC approach.
Fig. 1.6 Sequence of events Bipolar

during the electrodermal placement (SC measurement)
response Distal
phalanx
Medial
phalanx
Proximal
phalanx
C7
C6 Hypothenar
Thenar eminence
eminence Active electrode
C8 Dermatomal
Crease
distribution
Unipolar
placement
(SP measurement)
Reference
electrode
Abraded
site
Elbow
1.6.1 DC Source Front-End
As mentioned, DC recording of EDA is the most widely used method of mea-

surement. In the literature the DC recording has been recommended in several
publications [12, 22, 34] and there is no additional need to report further details
here. The most easy-to-construct circuit for exosomatic DC recording which can
be used in combination with a high-quality biosignal amplifier is shown in Fig. 1.7.
Generally, an integrated circuit is used to obtain a low constant voltage of 0.5 V
as specified in the guidelines (i.e., AC1 in Fig. 1.7). The current is converted to a
voltage by means of a current-to-voltage operational amplifier (inverting amplifier)
which is negative and proportional to the skin conductance of the subjects. This
output voltage is further amplified and inverted by means of another circuitry (i.e.,
AC2 in Fig. 1.7) and then usually connected to analog-to-digital converter.
Electrodes
vo AC2
vDC
0.5 V AC1 +
Fig. 1.7 Simplified circuit for the acquisition of the DC electrodermal activity
1.6.2 AC Source Front-End
A device for AC recording of EDA was proposed by Boucsein et al. [38], pointing
out that this method is preferred for medical applications to prevent nonlinearities
that may result from uncontrolled current densities when using a constant effective
voltage source.
With this method both impedance and phase angle are obtained as analog output
signals. Specifically they used an oscillator capable of generating a sine wave with
a continuously adjustable voltage between 1 Hz and 1 kHz. The sine wave was
converted by a voltage-to-current circuit into a constant current in the range between
0 and 10 mA of peak value, which is delivered to the subject.
The terminal voltage from the subject’s skin was preprocessed in two different
ways in order to compute both the impedance value and the phase angle. After a
pre-amplification step, the signal was amplified and rectified. Then, the impedance
was calculated after further low-pass filtering stage with a cutoff frequency of either
0.1 or 1 Hz and sent to an digital display. For calculating the phase angle, the pre-
amplified signal was multiplied in a phase sensitive detector with the oscillator
signal—which was phase shifted with the possibility of adjusting the phase angle
continuously—acting as a zero-offset for the phase signal. The output signal of the
phase sensitive detector was also rectified, low-pass filtered with the same frequency
limit as the impedance signal, and delivered to an other digital display.
A block diagram of the AC circuit for EDA recording is shown in Fig. 1.8.
calibration
100 100
4.73 nF output (2)
kW kW
phase angle[deg]
CAL 1 CAL 2
sub- digital
pre- phase low-pass
rectifier display
recording amplifier detector filter phase angle
.1 Hz
1 Hz
constant
current
source
peak- output (1)
current
0-10mA amount of
impedance
ject digital
phase voltage low-pass
rectifier offset display
shifter amplifier filter voltage
.1 Hz
0-100° 1 mV 1 Hz
10 mV
sinus 0° 100 mV
oscilator 90°
180° 1V
1Hz-1kHz 270°
x0... 10 10 v
1 kHz
10 Hz
1 Hz
Fig. 1.8 Block diagram of an analog front-end for simultaneous recording of impedance and phase
angle (courtesy of [38])
1.6.3 Remote DC Devices
Several devices and circuit designs for the acquisition of the exosomatic EDA can
be already found in the literature and on the market. Here, we discuss the most
commonly used among those which need to be connected to a remote computer
[52].
• ProComp. It is a multi-channel and multi-modality acquisition system used by
a number of researchers to record skin conductance signal in addition to other
physiological data used in clinical observation and biofeedback (e.g., EEG, ECG,
EMG . . . ) [53] as well as other derivated measures such as heart rate, blood
volume pulse, respiration, goniometry, force. It has 8 protected pin sensor inputs;
2 channels that read data at 2048 samples/second, and 6 channels that read it
at 256 samples/second. The connection between the remote computer and the
ProComp system is performed by means of a fiber-optic cable.
• BioPac. The BioPac MP [54] system is a data-acquisition board that can be
connected to a remote computer by USB or Ethernet cable. It is widely used in the
scientific community and also in this book has been used in several experimental
studies (see Chaps. 3 and 5). The BioPac system is able to collect a large variety
of physiological multi-modal data, including skin conductance, with adjustable
sampling frequency. Moreover, it is possible to set a digital preprocessing stage
including variable gain and different kinds of filter.
• Biosemi. Biosemi is a product intended to be used for research applications

only [55], and in particular it is a widely used device in neuroscience. Its common
use concerns the EEG signal acquisition, but with an external accessory, the
impedance of the skin can be measured. The EDA is acquired by means of two
passive electrodes to induce an oscillator signal synchronized with the sample
frequency. Because the BioSemi uses “Lock-in detection”, the stimulus-current
can be as low as 1uA. This acquisition method (i.e., the very low current and the
synchronized oscillator) allows that the other biosignals (i.e. ECG,EEG,EMG)
can be measured without any risk of to be corrupted by the EDA oscillator signal.
1.6.4 Wearable DC Devices
In the field of DC-devices, in the last decades, many wearable devices have been
developed in order to have the possibility of acquiring the EDA in daily-life
scenarios. They usually need a host computer to stream and memorize EDA data.
Most of them are designed as “a glove-like wearable device” in order to acquire the
EDA signal from the palm of the hand [56, 57]. An example is given by the so-
called Galvactivator [56] developed by the MIT Media Lab group. It is “a glove-like
wearable device” that senses the wearer’s skin conductivity and maps its values to a
bright LED display”. The galvactivator device also provides a data port from which
an analog to digital converter can sample. The sensor is comfortable, but requires
that the wearer be cabled to a host computer to transmit EDA data.
A commercial wearable DC device is the Brainquiry [58]. As a maker of “neuro-
feedback, biofeedback and psychophysiological measuring equipment”, Brainquiry
sells a compact skin conductance sensor which uses Bluetooth to communicate with
a host computer. However, little information is provided by the manufacturer about
the proprietary design of the biofeedback amplifier.
Finally, as mentioned before, in one of our previous study [57], a novel EDA
glove based on textile-integrated electrodes has been developed. The novelty
consisted mainly in the use of integrated textile electrodes placed at the fingertips.
The system is able to acquire and process the DC skin conductance in order
to discriminate affective states. The textile electrodes have been compared with
Ag/AgCl electrodes demonstrating comparable performance. More specifically,
reported results on electrode characterization, performed by means of the voltage-
current characteristics, and its electric impedance showed that textile electrode
achieves a good electrical and thermal coupling with biological site. Moreover, a
dedicated experiment where 35 subjects were enrolled and aiming at discriminating
different affective states using only EDA was designed and realized. A new set
of features extracted from non-linear methods was used, improving remarkably
successful recognition rates. Results were, indeed, very satisfactory and promising
in the field of affective computing.
Fig. 1.9 Sensorized glove

for the acquisition of the EDA
1.6.5 Multi-Frequency Sensorized Glove
As already anticipated, here we briefly preset a new wearable acquisition device,

which is able to acquire EDA data both with AC and DC sources: the multi-
frequency sensorized glove. The EDA signal is acquired by a glove with integrated
textile electrodes placed at the fingertips (Fig. 1.9). Moreover, the use of a wearable
textile system exhibits several advantages in terms of portability and usability for
long-term monitoring, and gives minimal constraints. This latter characteristics is
very significant when the system is used in an ecological environment.
The analog front-end of the designed electronics, which is responsible for
measuring the DC and AC exosomatic EDA, is based on a variable-gain current-to-
voltage operational amplifier. The electric current injected into the skin is variable
and programmable (from 0 to 1 kHz), and for this purpose we used the AD9833
provided by Analog Device [59]. This chip is a low power, programmable waveform
generator, which is used to switch up the frequency of the skin electrical stimulation
among 0 (i.e. DC), 10, 100 and 1000 Hz [40]. Moreover, a low-pass filter (cutoff
frequency of 3 Hz) and a further amplification stage were applied to the raw EDA
data before the successive digitalization step.
The preprocessed EDA signal was digitally converted thanks to the 12-bit analog-
to-digital converter built in the MSP430 microcontroller Fig. 1.10. The MSP430fxx
family of microcontroller is designed for low cost and, specifically, low power
consumption embedded applications. It is a very popular choice especially in
wireless networking systems and it is built around a 16 bit RISC CPU. In our
prototype, we used the MSP430x6xx Series, which are able to run up to 25 MHz,
have up to 512 KB flash memory and up to 66 KB RAM. Moreover, this series
includes an innovative power management module for optimal power consumption
and an integrated USB [60].
Fig. 1.10 Block scheme of the electronic circuit
Moreover, wireless communication was implemented by an Xbee module con-

nected to the USART of the MSP430. Specifically, it was used to exchange
data between the transceiver and a dedicated multi-platform software application.
Finally, a lithium-polymer battery with a voltage of 3.7 V and a capacity of around
750 mAh was chosen as power supply [39]. An external circuit was developed to
support the rechargeable battery through a USB port. Finally, a voltage regulator
is responsible for supplying 3.0 volts from the battery to all the components of the
device.
Chapter 2
Modeling for the Analysis of the EDA
As described in the previous chapter, EDA broadly refers to any alterations in the
electrical properties of the skin. The most frequently used measure of EDA is the
SC. The SC signal can be decomposed in two components, tonic and phasic, which
have different time scales and relationships to exogeneous stimuli. Tonic phenomena
include slow drifts of the baseline skin conductance level (SCL) and spontaneous
fluctuations (SF) in SC [15]. The phasic component, i.e., the skin conductance
response (SCR), reflects the short-time response to the stimulus. The typical shape
of SCR is comprised of a relatively rapid rise from the conductance level followed
by a slower, asymptotic exponential decay back to the baseline.
When the inter-stimulus interval (ISI), i.e., the temporal gap between two
consecutive stimuli, is shorter than the recovery time of the first response, the
two SCRs overlap. This occurrence is observed in many experimental paradigms,
particularly in cognitive neuroscience where common values of ISI (1–2 s) are
generally shorter than the recommended minimum ISI to avoid such an overlap,
which is around 10–20 s [17, 61]. The overlap issue is probably the main limitation
in the treatment of the decomposition of SC into its phasic and tonic components.
Despite the wide use of EDA measurements and related research, the generation of
SCR via skin sympathetic nerve fibres is still unknown.
2.1 Mathematical Models of the EDA: An Overview
In the past two decades, several mathematical solutions have been developed to
decompose the phasic signal into individual SCRs associated with each stimulus,
even during short ISI experimental paradigms, and to model how ANS activity
(and, in particular, the sudomotor nerve activity) causes SCRs. This process allows
estimation of ANS activity with potentially better time resolution than using the
raw SCR signal. Many of the early methods, whose primary aim was to overcome

20 2 Modeling for the Analysis of the EDA
the overlap issue, required visual inspection and introduced subjective elements
into the analysis. For example, Barry et al. [62] attempted to correct the baseline
by subtracting each SCR from an extension of the preceding SCR using graphical
tools. Lim et al. [63] proposed a model based on a response function made of 4–
8 parameters optimized for each single response to obtain a response-by-response
variation in SCR shape. This method also required visual inspection to select the
best model setting.
Further assumptions have been related to the description of the peripheral ner-
vous system as a linear time-invariant (LTI) system [64]. In addition to decomposing
the phasic signal into individual SCRs, these models often attempt to estimate the
ANS activity by searching for the most likely input signal which could explain the
observed output (the measured SC). The first LTI model for EDA analysis was
presented by Alexander et al. [65]. Their method allowed the estimation of the
sudomotor nerve activity (SMNA) using a model where the SC is the result of a
convolution between discrete bursting episodes of the SMNA and a biexponential
impulse response function (IRF) assumed to be known a priori and time invariant.
Benedek and Kaernbach criticized some aspects of Alexander’s model and
developed two new models in which the LTI assumption was modified to take into
account the variability in SCR shape. In addition, they considered mathematical
constrains to significantly improve the reliability of the physiological modeling.
These models are known as the non-negative deconvolution model [16] and the
continuous deconvolution model [1]. Both models split the SMNA into two parts,
one describing the actual phasic activity and the other representing EDA variations
of different origins (e.g., noise). Both models assume a pharmacokinetic model of
the dynamic law of diffusion of sweat. They adopted a biexponential IRF, called the
Bateman function. Although observation noise is not formally modeled in any of
these methods [1, 16, 65], all of three assume its existence. They estimate a noisy
SMNA and then recover a filtered phasic component using a low-pass filter and a
subsequent heuristic and prefixed peak-detection scheme.
Recently, Bach et al. presented the SCRalyze toolbox (now incorporated into
PsPM and available online at: pspm.sourceforge.net), which comprises several
models that assume a linear time-invariant system [66]. Note that Bach’s model also
imposes mathematical constrains to improve the physiological mimicking. These
models and that of Alexander et al. use a heuristic IRF whose parameters have been
optimized on large datasets. SCRalyze algorithms try to estimate the model input
(SMNA) or parameters that best explain the observed SC data based on optimization
methods. Moreover, they include a noise term, which also accounts for possible
violations of the assumption of time invariance.
More recently, Chaspari et al. [67] proposed a sparse representation of EDA
but their use of overcomplete dictionaries leads to a non-convex problem with no
guarantee of finding the globally optimal solution. Since an great variety of practical
problems can be cast in the form of a convex optimization problem, mathematical
optimization has become an important tool in many disciplines and the list of its
applications is steadily growing [68].
2.2 EDA Analysis 21
In this book, we report on a recently proposed method called CvxEDA [69] to

estimate the ANS activity from the EDA using a convex optimization approach. The
model is grounded on Bayesian statistics and a simple yet physiologically sound
representation of the observed SC as the sum of three components: a slow tonic
component; the output of the convolution between an IRF and a sparse (compact,
bursty) non-negative SMNA phasic driver; and an additive noise term. Of note, the
IRF, which is related to the phasic component, is modeled as output of an Infinite
Impulse Response (IIR) system.
2.2 EDA Analysis
EDA is a widely used measurement in, e.g, the psychophysiology research field.
In this context, one of the main goal is to infer on perceptual affective states
from physiological, peripheral, may be non-invasive measurements such as SCR.
However, sometimes psycho-physiological processes (e.g., sympathetic arousal) can
have higher time resolution than the observed variables (e.g., SCRs). This shortcom-
ing dramatically impact on the use of conventional analysis, which does not employ
any a-priori modeling. Consequently, recent years have seen an increased interest
in the design of causal models, aiming to estimating unobservable processes from
observable ones such that inference can be drawn from the unobservable variable
directly [64, 66].
In the following text, we take into account two different approaches: the
conventional and the model-based approach.
2.2.1 Conventional Analysis
The aim of conventional analysis is to extract features from the observable variable,
in this case the SC signal, that can closely represent a psychological central state.
A conventional data analysis algorithm can be described as follows: (1) data filtering
to reduce the observation noise, (2) definition of a time response window in order to
identify only the peaks that are stimulus-evoked, and (3) definition of some criteria
to detect peaks within this window (e.g., a threshold beyond which the SC peak can
not be considered significant, see Fig. 2.1). After the identification of peaks, feature
extraction can be performed. These features are usually defined through qualitative
or semi-quantitative models.
Peak
Feature
Data filtering Segmentation detection
extraction
criteria
Fig. 2.1 General diagram of a conventional analysis procedure
2.2.2 Model-Based Approach
The model-based approach describes how the observable process, i.e., EDA, is
generated by the central process, (e.g., sympathetic arousal) using mathematical
equations that formulate psychophysiological assumptions. Therefore, the model
predicts the SC time series, and the independent variable consists in the unobserv-
able psychological status. However, in the analysis of experimental data we have to
consider the opposite situation: we know the observed SC data but not the central
state, and we try to estimate the time series of the central process that generated
these SC data. To this extent, the forward model has to be turned backwards,
to find the relation between SC and central state. In statistics, this mathematical
process is often termed “model inversion”. In conclusion, both conventional and
model-based analysis try to infer the CNS state from the peripheral signals, but the
difference is that model-based methods use a more stringent mathematical language
and computational methods to do so, while the general aim is the same [64].
Usually, the model-based methods distinguish two steps in the relationship
between the central state and the SC data: the neural model that specifies how the
central state, in terms of event-related or spontaneous sympathetic arousal, elicits
SMNA, and the peripheral model, from the SMNA to the SC, that specifies how
SMNA generates SC usually in the form:
SC D SMNA IRF
2.3 CDA: Continuous Deconvolution Analysis 23
where is the convolution operator, and IRF is a skin conductance impulse response
function. This model is, in a basic version, deterministic. Through this approxima-
tion, the SC time series is only influenced by SMNA and not by other confounding
factors such as noise (note that the deterministic deconvolution enhances noise [16]).
Different model inversions schemes treat this problem differently.
2.2.2.1 Model Evaluation
The evaluation of a model and the comparison between two or more methods is
hard to be performed. Each method returns an output that can be an index or the
central state, but we have to measure which method estimates the central state at
best. A possible evaluation process uses an experimental paradigm in order to create
two central states that are known to be different. Thus, the method can be evaluated
investigating the ability to detect this difference. This is achieved using a statistical
approach on the features extracted that quantify the central activation.
In this book, we first present a conventional modeling approach, the Continuous
Deconvolution Analysis (CDA) [1], and a recently proposed model proposed by
Greco et al. based on rigorous mathematical definitions: the convex optimization
approach (cvxEDA). In the next chapters, performance of this model will be
empirically evaluated and compared.
2.3 CDA: Continuous Deconvolution Analysis
The CDA is a method, proposed by Benedek et al., to decompose the skin

conductance signal into its tonic and phasic driver data. It is based on an explicit
biophysical model, and its parameters are optimised for each individual dataset. The
decomposition process comprised in three different steps: a preprocessing phase, in
which the signal is filtered to reduce the noise, a deconvolution process in order
to obtain the phasic and tonic driver, and an optimization stage to improve the
estimation of the parameters of the impulse response function. The decomposition
process can be performed by means of the toolbox Ledalab. software package for
MATLAB [70], which is available online (www.ledalab.de).
2.3.1 Preprocessing
In the preprocessing stage, the detection of movement artifacts was carried out
by visual inspection. Artifact-free signals exclusively were taken into account for
further analysis. In order to limit the frequency bandwidth of the EDA signal, it was
filtered with a low pass zero-phase forward and reverse digital filter [71, 72] with a
cutoff frequency of 2 Hz, having Buttworth approximation.
2.3.2 EDA Deconvolution Analysis
EDA is produced by changes in the skin conductivity as major effect of the sweat
gland activity. Specifically, sweat is released to the sweat duct, passes to the stratum
corneum, and finally is brought out of the skin. Accordingly, the dynamics of the
variation of concentration of sweat in the stratum corneum can be represented by
a two-compartment pharmacokinetic model in which the sweat concentration is
assumed to change only by diffusion [14, 73]. The first compartment represents
the sweat duct and the second compartment the stratum corneum. Being the two
compartments different in dimension (i.e. the stratum corneum is much larger
than the sweat duct), the diffusion can be considered as a one way-diffusion.
Solving the two coupled first-order differential equations of each compartment, the
solution is the Impulse Response Function IRF.t/ which is also known as Bateman
function [74]:
t t
IRF.t/ D .e 1 e 2 / u.t/ (2.1)
The Bateman function is characterized by a steep onset and a slow recovery. The
steepness of onset and recovery is determined by the time constants 1 and 2 .
EDA can be divided into tonic (SCL: Skin Conductance Level) and phasic
components (SCR: Skin Conductance Response). The tonic electrodermal com-
ponent represents the baseline level of the signal whereas the phasic component
indicates a direct response to a specific stimulus. However, there are often phasic
parts of EDA which cannot be related to any specific stimulus, and hence, they are
called spontaneous or nonspecific SCRs [23]. When the time interval between two
consecutive stimuli is shorter than the recovery period of SCR, the stimuli responses
in the SCR are overlapped. In this case, the typical shape of the SCR is lost and this
could be one of the main issue for the extraction of the correct information from
the electrodermal signal. In order to overcome this issue, the EDA signal process is
modeled as a convolution process between the SudoMotor Nerve Activity (SMNA),
as part of the sympathetic nervous system, and IRF [1] under the hypothesis that
EDA is controlled by SMNA resulting in a sequence of distinct impulses which
regulate the eccrine sweat glands dynamics (see Fig. 2.2).
Formally, it is possible to write:
Pre-processing Decomposition
Raw signal • Segmentation (Deconvolution) SMNA
(EDA) • Filtering • IRF=(exp(-t/τ1)-exp(-t/τ2))u(t)
Fig. 2.2 Electrodermal acquisition and decomposition process. The EDA is filtered to reduce the
noise and then decomposed in tonic and phasic components by means of a deconvolution with an
impulse response function (IRF) called Bateman function
2.4 CvxEDA: A Convex Optimization Approach to Electrodermal Activity. . . 25
EDA D SMNA ˝ IRF (2.2)
where SMNA D .DRIVERtonic C DRIVERphasic /. In the Eq. (2.2), SMNA is unknown

and is evaluated by deconvolving the EDA signal with the IRF. To decompose the
obtained SMNA signal into the DRIVERtonic and DRIVERphasic components, several
algorithmic steps should be taken into account. A smoothing Gauss window of
200 ms is applied to SMNA, followed by a peak detection algorithm in order to
find the peaks over a threshold of 0:2 S. All the points below the threshold were
interpolated with a cubic spline fitting method giving the DRIVERtonic . More details
can be found in [1]. Finally, the DRIVERphasic component, instead, is computed by
subtracting the previously estimated DRIVERtonic from the SMNA (see Fig. 5.23),
under the hypothesis that tonic activity is observed in the absence of any phasic
activity [23].
Of note, the DRIVERphasic signal should have a zero baseline intermitted by
distinct peaks overcoming the issue of having overlapped SCRs.
2.3.3 Optimization
Starting from fixed values, the parameter set of the IRF (i.e. 1 and 2 ) was optimized
according to criteria evaluating the quality of the model, through the minimization of
a specific cost function given by the sum of the number of points of the DRIVERphasic
component that have negative value and the number of points above a predefined
threshold (equal to 5 % of the maximum of DRIVERphasic ). This procedure aims at
having a signal with a zero baseline a peaks as distinguishable as possible. More
details can be again found in [1].
2.4 CvxEDA: A Convex Optimization Approach

to Electrodermal Activity Processing
CvxEDA has been inspired by some aspects of the CDA model, such as the
Batman function. This novel algorithm is based on three main concepts: maximum
a posteriori probability, convex optimization, and sparsity.
2.4.1 Convex Optimization
A set K 2 Rn is convex if:
x C .1 /y 2 K (2.3)
8 x; y 2 K and 2 Œ0I 1. A function is convex if:
f .x C .1 /y/ f .x/ C .1 /f .y/ (2.4)
8 x; y 2 K and 2 Œ0I 1. The meaning of the inequality (2.4) is that, for any
two points x and y in the domain of the function, the segment between .xI f .x// and
.yI f .y// lies above the graph of the function. Equivalently, we can define a convex
function as a function whose epigraph is a convex set [68].
Considering a standard optimization problem:
minimize f0 .x/
(2.5)
subj. to fi .x/ 0 i D 1; : : : ; m ;
the optimal choice is the one minimizing the objective function f0 .x/, which
represents the cost of choosing x, while simultaneously satisfying the constraints
fi .x/ 0. An optimization problem is convex when both the objective and the
constraint functions are convex. In the context of mathematical optimization, the
most important consequence of convexity is that necessary conditions for local
optimality are also sufficient for global optimality. Moreover, important categories
of convex optimization problems can be solved efficiently (this is rarely the case for
general nonconvex problems).
A special subclass of convex optimization problems is represented by least-
square problems where the goal is the unconstrained minimization of a quadratic
objective function kAx bk22 . For this class of problems—frequently arising in
regression analysis, parameter estimation and data fitting methods [68]—an ana-
lytical solution exists. An important statistical interpretation is that the least-square
solution coincides with the maximum likelihood estimation in the case of a linear
model corrupted by additive Gaussian noise. Regularization, e.g. adding a norm of
the optimization variable x as an extra term to the cost function, can be applied to
least-squares problems to prevent overfitting (L2 -norm) or to favour sparse solutions
(L1 -norm). While in the former case an analytical solution exists, in the case of the
L1 -regularization the problem can be cast as a quadratic program (QP), i.e. a convex
problem with quadratic cost function and affine constraints:
1
minimize xtrP x C qtrx C r
2 (2.6)
subj. to Hx g 0 and Ux v D 0 :
2.4.2 Model Assumptions
In this model, the EDA generation process based on the following assumptions:
A1) SCRs are preceded by bursts from the sudomotor nerves controlling the sweat
glands. These bursts are temporally discrete episodes [75, 76], i.e. SCRs are
generated by a neural signal that is sparse and non-negative because of the
nature of a nerve activity.
A2) The relationship between the number of sweat glands recruited and the
amplitude of a firing burst is linear [76]. Moreover, the output response of
the system depends only on the instant where the nerve input is applied. Stated
otherwise, the timecourse of a single SCR induced by a neural burst is not
influenced by previous ones, even when their SCRs overlap [77]. In the light
of these considerations it is reasonable to characterize the system as linear
time-invariant.
A3) The sweat diffusion process has a subject-specific impulse response function
(IRF) which is relatively stable for all SCRs from the same subject [1].
A4) This phasic activity is superimposed to a slowly varying tonic activity with
spectrum below 0:05 Hz [78], i.e. whose information content can be repre-
sented by samples spaced every 10 s (e.g., by 10-s averages in [15]).
2.4.3 Observation Model
A given N-sample long SC signal (y) is modeled as the sum of a tonic (t) and a
phasic (r) component plus an additive noise term ():
y D r C t C ; (2.7)
where y, t, r, and are N-long column vectors. The noise term is an iid (inde-
pendent and identically distributed) sequence of zero-average Gaussian random
variables with variance 2 , representing measurement and modelling errors.
The tonic component is represented as the sum of cubic B-spline functions with
equally-spaced knots every 10 s (assumption A4), an offset and a linear trend term:
t D B` C Cd; (2.8)
where B is a tall matrix whose columns are cubic B-spline basis functions, ` is the
vector of spline coefficients, C is a N 2 matrix with Ci;1 D 1 and Ci;2 D i=N, d is
a 21 vector with the offset and slope coefficients for the linear trend.
Within r, the shape of a single phasic response (under assumptions A2 and A3)
is modelled using a biexponential impulse response function, called the Bateman
function:

h. / D .e 0 e 1 / u. /; (2.9)
where 0 and 1 are, respectively, the slow and fast time constants while u. / is the
unitary step function. The Bateman function is the output of a bi-compartmental
pharmacokinetic model representing the diffusion of the sweat through the gland
ducts [73]. The Laplace transform of (2.9) is simply:
˚ 1 1
L h. / D 1
; (2.10)
s C 0 s C 11
where 01 and 11 are the poles of this second-order LTI system. Its discrete-
time approximation, obtained using central differencing (bilinear transform) s D
2 z1
ı zC1
with sampling interval ı, is the following ARMA model:
2
1 C z1
H.z/ D
C z1 C z2
D .11 ı C 2/.01 ı C 2/=.11 ı 2 01 ı 2 / (2.11)
D .2 11 01 ı 2 8/=.11 ı 2 01 ı 2 /
D .11 ı 2/.01 ı 2/=.11 ı 2 01 ı 2 /:
The ARMA cascade can be represented in matrix form as:
q D A1 p; r D M q; (2.12)
where: p represents the sudomotor nerve activity; q is an auxiliary variable that will
be used to find p indirectly; M is a tridiagonal matrix with elements Mi;i D Mi;i2 D
1, Mi;i1 D 2, 3 i N; and A is a tridiagonal matrix with elements Ai;i D ,
Ai;i1 D , Ai;i2 D , 3 i N.
Finally, the observation model (2.7) can be written as:
y D Mq C B` C Cd C : (2.13)
2.4.4 Maximum a Posteriori Estimation
Given the observation model (2.13), the goal is to identify the maximum a posteriori
(MAP) spike train (p) and tonic component (t) parametrized by Œq; `; d, for the
measured SC signal (y):
Œq; `; d D arg maxq;`;d PŒq; `; d j y: (2.14)
Assuming independence between q, ` and d (i.e. between the phasic activity, the
slowly varying tonic component and the drift) and applying Bayes’ theorem, we
obtain:
PŒq; `; d j y / PŒy j q; `; d PŒq PŒ` PŒd; (2.15)

where PŒy j q; `; d is the likelihood of observing a specific SC time series given

the parameters of the model, while PŒq, PŒ` and PŒd are the prior probabilities of
the parameters. In (2.15), we omitted the evidence of the SC data PŒy since it plays
no role in the optimization. Unlike other approaches in the literature, our model
relies exclusively on the presence and definition of the priors in (2.15)—which we
are about to describe in detail—to impose physiologically sound constraints on the
signals to be estimated. As a result, the method does not require pre-processing of
the observed SC signal (e.g., bandpass filtering) nor post-processing of the inferred
phasic and tonic components (e.g., to deal with negative neural activations).
To model the sudomotor nerve activity (p) representing the input (A1) to the LTI
system, the simplest first order description of spike trains is used [79], i.e. a Poisson
distribution:
pi Pois.ı/; (2.16)
where ı is the expected firing rate per bin, i.e. is the average number of spikes
per unit time. To keep the analysis tractable, the Poisson distribution is replaced with
an exponential distribution of the same mean [79]. In this way the constraint pi 2 N
can be relaxed to pi 0. Finally, since p and q are related by (2.12), the prior PŒq
becomes:
YN 1 1 pi YN 1 .Aq/i
PŒq D e ı / e ı : (2.17)
iD1 ı iD1
Concerning the tonic component, the authors make use of assumption A4 and
consider a uniform frequency spectrum in the band 0 0:05 Hz. Because equally-
spaced knots every
D 10 s are used, the sampling frequency is exactly twice the
upper band limit and the elements of the vector ` can be assumed iid. In particular,
a normal distribution is adopted for the amplitude at each knot `i N .0; `2 /. As a
result the prior PŒ` is:
YQ
1 1 `2i
PŒ` D p exp 2 ; (2.18)
iD1 2 ` 2 `
where Q is the number of knots (approximately Nı=

). Finally, for the drift
coefficients d an uninformative priors is assumed and drop PŒd altogether from
further analysis.
The likelihood term follows immediately from (2.13) and from the error model
N .0; 2 /:
!
YN 1 .Mq C B` C Cd y/2i
PŒy j q; `; d D p exp : (2.19)
iD1 2 2 2
Replacing (2.17), (2.18) and (2.19) in (2.15) and taking the logarithm:
1 XN
ln PŒq; `; d j y D .Mq C B` C Cd y/2i
2 2 iD1
1 XN 1 XQ 2
.Aq/i 2 ` C const; (2.20)
ı iD1 2` iD1 i
with .Aq/i 0. Maximizing (2.20) yields the MAP solution to (2.14). After
multiplying by 2 and substituting ˛ D 2 =.ı/ and D 2=`2 , (2.20) is rewritten
as a constrained minimization problem in matrix form to obtain a more compact
notation. This optimization problem, that is termed cvxEDA, represents the core of
the algorithm presented in this manuscript:
1
minimize kMq C B` C Cd yk22 C˛ kAqk1 C k`k22
2 2 (2.21)
subj. to Aq 0:
After some matrix algebra, this optimization problem can be re-written in the
standard QP form and solved efficiently using one of the many sparse-QP solvers
available. After finding the optimal Œq; `; d, the tonic component t can be derived
from (2.8) while the sudomotor nerve activity driving the phasic component can be
easily found as p D Aq.
Although solving (2.21) is strictly equivalent to maximizing (2.20), the former
has a different interpretation. In the optimization problem, the objective function
to be minimized is a quadratic measure of misfit between the predicted and the
observed data. Prior knowledge is accounted for by means of additive regularizing
terms. For example, the spiking nature of the driving input (assumption A1) is
enforced by means of the l1 -norm penalization which is an effective way to sparsify
a signal while maintaining convexity [80–82]. Smoothness of the tonic curve
(assumption A4) is enforced by the choice of the basis (B) and through the l2 -
norm penalization of the spline coefficients. The two parameters ˛ and control
the strength of the penalty for the phasic and tonic components, respectively. A
large ˛ (stronger l1 regularization of p) yields a sparser estimate with most noise-
induced spurious spikes suppressed but also more signal distortion (i.e. attenuation
of genuine activations). Conversely, a small ˛ produces a less distorted but noisier
solution. Concerning , higher values mean a stronger penalization of `, i.e. a
smoother tonic curve.
Of note, CvxEDA algorithm is implemented in Matlab language and the soft-
ware is available online (www.mathworks.com/matlabcentral/fileexchange/53326-
cvxeda).
2.5 Feature Extraction 31
2.5 Feature Extraction
Regardless of the model or the type of analysis used to perform the decomposition
of the EDA signal, several features are extracted from tonic and phasic signals in
order to assess the sympathetic system activity.
2.5.1 Time Domain
Typically, time-domain features of EDA are widely used to quantify the overall
activation of the ANS. Features extracted from the phasic signal are usually calcu-
lated into time windows of 5 s after the onset of the external stimulus (according
to the knowledge that SCRs arise within 1–5 s after the stimulus onset [16, 83]).
Features extracted from the tonic component express the sympathetic tone and are
often computed within time windows of 20 s, since the upper cut-off frequency of
the tonic component is about 0:05 Hz [84]. Moreover, within the group of tonic
measurements, nonspecific skin conductance responses are included, by definition.
They are often characterized by their frequency and mean amplitude within the time
window analysis.
In Table 2.1, the features set is summarized along with the corresponding
description.
Table 2.1 List of the features extracted from the EDA phasic and tonic
components
Feature Description
nSCR Number of significant SCRs within
the time response windows (WTRW) of 5 s
MAX-Tonic Maximum value of the tonic curve
within the time window
MAX-Phasic Maximum value of the phasic curve WTRW
AUC-Tonic Area under the tonic curve over time
AUC-Phasic Area under the phasic curve WTRW
Mean-Tonic Mean value of the tonic component over time
Mean-Phasic Mean value of the phasic component WTRW
STD-Tonic Standard deviation of the tonic component
STD-Phasic Standard deviation of the phasic component WTRW
NsSCR freq Frequency of the NsSCRs
Mean-NsSCR Mean value of the NsSCRs component over time
2.5.2 Frequency Domain
In addition to the features defined in the time domain, recently Posada-Quintero et

al. [85] proposed an analysis of the Power Spectral Density (PSD) of the EDA signal
in order to assess the sympathetic nervous system activity.
The effective assessment of the sympathetic dynamics, in fact, can have impor-
tant diagnostic impact and is one of the major fields of interest in cardiovascular
research [86]. The most common way to assess the ANS dynamics is to compute
the PSD of the HRV [87]. The HRV spectrum can be considered as characterized
by three components: the high-frequency (HF) bandwidth, from 0.15 to 0.4 Hz, that
is known to be influenced only by the parasympathetic nervous system activity;
low frequency (LF) bandwidth, from 0.045 to 0.15 Hz, which is influenced by both
the sympathetic and parasympathetic nervous system activities, and the very low
frequency (VLF) from 0.0033 to 0.04 Hz. Therefore the ratio between LF and HF,
which was typically used to assess the ANS balance, is not fully accepted as an
accurate measure of the balance between sympathetic and parasympathetic systems,
since the LF band also contains parasympathetic dynamics.
Thus, since EDA is directly controlled only by the sympathetic branch of the
ANS, Posada-Quintero et al. examined if a similar association of LF components of
HRV to sympathetic function also exists with the sudomotor function as measured
by EDA. In fact, in this way the main disadvantage of LF domain analysis of HRV
could be overcame because EDA is not influenced by the parasympathetic nervous
system activity.
In their study, Posada-Quintero et al., considered only tonic and NsSCRs, since
they were mainly interested in tonic stress responses, and excluded from the
analysis the evoked SCRs. The time-domain parameters of the tonic components,
such as the tonic mean value as well as the NsSCRs frequency (see Fig. 2.3), are
strictly related to the sympathetic activity but these indices are highly variable
among subjects [88]. Therefore, they have calculated the PSD of the EDA (see
Fig. 2.3) investigating the assumption that, if EDA can represent the cardiac and
peripheral sympathetic nervous systems dynamics, the spectral power should be
largely present in the low frequency band (0.04–0.15 Hz). Actually, after three
experimental paradigms concerning orthostatic, physical and cognitive stress, they
concluded that the frequency response of the sympathetic activities represented in
the EDA signal can be defined to be within 0.045–0.25 Hz. In addition, analyzing
EDA in the frequency domain, the PSD can lead to less inter-subject variability as
compared to the tonic features defined in the time domain (which are known to be
reliable indices of the overall sympathetic activation), due to the inherent filtering
properties of the frequency domain transformations. Finally, as stated above, the
EDA represents only the sympathetic nervous activities, whereas the LF power
of the HRV comprises both parasympathetic and sympathetic activities. Therefore,
thanks to the analysis in the frequency domain, it is possible to obtain a marker of
the sympathetic activity, which can be considered more reliable and sensitive than
the LF power of HRV.
2.5 Feature Extraction 33
Fig. 2.3 Diagram of the signal processing procedure to extract EDA and HRV parameters.
Courteously from [85]
Chapter 3
Evaluation of CDA and CvxEDA Models
There is no universally accepted experimental protocol for the validation of EDA

analysis algorithms. A characterization in terms of sensitivity and specificity, as
typically done in a pattern recognition framework, is not directly applicable in
this context since there is no one-to-one correspondence between external stimuli
supposed to elicit ANS responses and skin conductance responses (see [64] for
a discussion). Unless the sympathetic nerve activity is also recorded through
microneurography, failure to detect a phasic SC response after the occurrence
of an experimental stimulus may be equally ascribed to a low sensitivity of the
algorithm under study or, alternatively, to the inability of the stimulus to consistently
elicit a phasic response. Similarly, detection of phasic activity in the absence of
stimulation may be caused by electrodermal changes that are not stimulus-elicited
but spontaneous and non-specific, possibly a result of muscular contractions or
respiratory irregularities [15]. We tested the new model taking into account evalua-
tion procedures reported in the literature. Moreover, performances of cvxEDA were
also compared with those obtained from the continuous deconvolution analysis, as
implemented in the Ledalab software [1]. We first validated the model’s ability
of estimating phasic and tonic components explaining the observed SC through
synthetic and experimental data (see “Experiment 1” below). In the latter case,
we used a forced maximal expiration task, that a previous study [89] has shown
to reliably induce a sympathetic activation. A further validation was performed to
investigate the predictive power of features derived from cvxEDA model in inferring
central (mental/emotional) states (“Experiment 2”).

36 3 Evaluation of CDA and CvxEDA Models
3.1 Synthetic Data
Each synthetic SC time series lasted T D 90 s and was generated as the sum of three
terms: the first one, representing the phasic component of the EDA, was obtained
as the result of a convolution between a synthetic SMNA and a biexponential IRF
(1 D 0:7 s, 0 Unif.2:0; 4:0/ s); the second one was a slowly varying signal
representing the tonic component, obtained as a linear trend plus a sinusoid with
a period Tt Unif.45:0; 90:0/ s; the third term was an additive white Gaussian
noise (AWGN). The sudomotor nerve activity driving the phasic component was
simulated by placing 10 pulses of unit area (modelling neural bursts) at random
times with a minimum 1-s distance between them and from the two ends. To test the
ability of the two methods to recover partially overlapping SCRs in the presence of
noise, two sets of 100 time series were generated with different levels of signal-to-
noise ratio (SNR): 33 dB and 13 dB (defined as 10 log10 .a2 =N2 / where a is the foot
to peak amplitude of a single SCR and N2 is the AWGN variance).
3.2 Experiment 1: Maximal Expiration Task
In the first experiment, 15 healthy subjects (aged 18–35 years; 7 females) performed
a forced maximal expiration task [89], in which they were asked to breathe out with
the maximum possible intensity in order to trigger the ANS-mediated expiration
reflex. All subjects gave written informed consent prior to taking part in the study,
which was approved by the local Ethics Committee. A Biosemi Active II system
was used to acquire the SC signal and the respiratory effort (by means of a thoracic
respiration belt). The protocol started with the subjects breathing normally and
resting in front of a grey monitor for 3 min in order to record their baseline levels.
This was followed by three stimulus sessions in which subjects had to perform a
deep expiration whenever the colour of the screen background changed to black.
Each session consisted of six forced expirations with a variable ISI chosen randomly
among 4, 8 and 12 s. Consecutive sessions were separated by a 30-s recovery
interval.
This experimental paradigm was chosen to obtain SC signals in which the
presence of an autonomic response to the stimulus was as objective and reliable
as possible. In fact, previous studies have shown that the forced expiration protocol
is a valid method of evoking SCRs unaffected by emotional change with more stable
waveform patterns, less habituation and better reproducibility than other means
of stimulation (including electrical) [89]. In this way, the presence of at least one
SCR after each stimulus was ascertained, allowing determination whether the new
methodological approach was able to separate and identify each phasic response
even when stimuli were close to each other and their SCRs overlapped.
3.4 EDA Processing and Analysis 37
Table 3.1 Arousal rating of Session Arousal rating Arousal range Arousal level
IAPS images used
N 2:81 ˙ 0:24 2.42–3.22 VL
A1 3:58 ˙ 0:30 3.08–3.98 L
A2 4:60 ˙ 0:31 4.00–4.99 L-M
A3 5:55 ˙ 0:28 5.01–6.21 M-H
A4 6:50 ˙ 0:33 5.78–6.99 H
3.3 Experiment 2: Visual Affective Stimuli
In the second experiment, 15 healthy subjects (aged 22–26 years; 7 female) different
from the previous ones were stimulated by viewing affective images from the
official IAPS database [90] to assess algorithm’s predictive validity, i.e. its ability
to distinguish stimulations with different arousal content and provide meaningful
information about ANS activation. All subjects gave written informed consent
before taking part in the study, which was approved by the local Ethics Committee.
Subjects were comfortably seated in an acoustically insulated room watching the
slideshow on a computer screen while their SC was recorded using a BIOPAC
MP150 physiological acquisition system. The affective elicitation consisted of four
arousal sessions alternated with four neutral sessions: N, A1, N, A2, N, A3, N, A4;
where N sessions are sequences of 6 very low arousal (VL) images while Ai (with
1 i 4) are sets of 20 images eliciting increasing levels of arousal. Details about
arousal rating values are reported in Table 3.1. Arousal sessions were classified as
Low (L), Low–Medium (L–M), Medium–High (M–H) and High (H) according to
the IAPS score criteria. Each image was presented for 10 s.
3.4 EDA Processing and Analysis
For each dataset, the continuous deconvolution analysis and the convex-
optimization-based EDA model described in the previous chapter were applied
to each SC time series. As per assumption of linearity and time-invariance system, a
subject-specific IRF was considered for this study. Concerning the CDA algorithm
1 and 2 parameters were optimized according to criteria for reduction of the
number of points of the phasic driver component that have negative value and the
number of points above a predefined threshold. Concerning the cvxEDA model,
while 1 D 0:7 s was used for all subjects, the optimal 0 was determined on a per-
subject basis as the value 0 2 Œ2:0; 4:0 s that minimized the l2 -norm of the residual
after fitting the cvxEDA model. Fixed values ˛ D 0:4 and D 0:01, chosen during
previous exploratory tests on separate data, were employed throughout this analysis.
The accuracy of the algorithms on the synthetic dataset was assessed by
measuring its ability to recover the neural activations in the phasic driver from
noisy SC time series. For each time series, the set of occurrence times T e of pulses
with area exceeding a 0:5 threshold was compared to the set of times T s of the
impulses in the synthetic SMNA using an algorithm modelled after the AAMI/ANSI
EC38:1998 standard. Briefly, times in T e and times in T s were considered as
“matching” if they were within a match window of ˙0:15 s. Each impulse from
either signal could only match a single impulse from the other one. Times in T e not
matching any element of T s were considered false positives (FP) while times in T s
not matching any element of T e were considered false negatives (FN). Finally, the
performance of the algorithm was measured in terms of sensitivity, computed as the
fraction of matched elements of T s , and positive predictive value (PPV), computed
as the fraction of matched elements of T e .
In the respiratory stimulation dataset, the presence of an estimated burst of
SMNA activity was verified in each 5-s time window following a stimulus onset,
in order to prove the model’s ability to correctly detect real SCRs.
In the last study, to verify that the recovered components represented meaningful
information regarding ANS activity, we investigated whether the amplitude of the
phasic driver p increased in response to affective stimulation with increasing levels
of arousal, as previously reported in the literature [91, 92], [15, Chap. 2.2.2]. An
intersubject analysis compared the responses to the four arousal levels through a
non-parametric Page test [93, Chap. 7.2], under the alternative hypothesis of increas-
ing phasic responses with increasing levels of arousal (we used non-parametric
tests because the hypothesis of Gaussianity was rejected by a Kolmogorov–
Smirnov test, p < 0:05). In post-hoc analysis, each pair of arousal sessions was
compared using a one-tailed Wilcoxon signed-rank test with Bonferroni correction
to determine significant differences between arousal levels in the expected direction.
We computed the adjusted p-value, i.e. the original p-value multiplied by 6 (the
number of pairwise comparisons among 4 conditions), to allow direct comparison
to the standard significance levels (e.g., 0:05). In the following,
p we also report the
Z-scores (from which the measure of effect size ZN D Z= N can be computed,
where N D 15 is the sample size). Finally, the slow tonic component was analyzed
comparing mean values of each arousal session with the preceding neutral session,
using a one-tailed Wilcoxon signed-rank test.
3.5 Experimental Evaluation Results
For all EDA datasets analyzed, the cvxEDA model produced the expected results:
the SC data (Fig. 3.1a) was decomposed into two signals, a sparse component p
and a smooth component t, that we interpret as the activity of the sudomotor nerve
(Fig. 3.1b) and the tonic level (Fig. 3.1c).
3.5 Experimental Evaluation Results 39
2.0 (a)
y [n.u.]
1.0
0.0
–1.0
1.0
(b)
0.8
p [a.u.]
0.5
0.2
0.0
2.0 (c)
t [a.u.]
1.0
0.0
–1.0
0 100 200 300 400 500 600 700 800 900 1000
τ [s]
Fig. 3.1 Application of the cvxEDA decomposition procedure to the SC signal recorded during
the forced maximal expiration task for a representative subject. (a) Raw SC signal, Z-score
normalized. (b) Estimated sparse phasic driver component p. (c) Estimated slow tonic component t
3.5.1 Results on Synthetic Data
Application of the cvxEDA model to the synthetic dataset highlighted the sparsity
of the p term, as well as the smoothness of the tonic component, even with low SNR.
Qualitative visual inspection analysis of Fig. 3.2 was sufficient to determine that the
algorithm worked properly on these data. A quantitative proof was also pursued in
terms of detection performance while recovering the neural activations in the phasic
driver from the noisy SC time series. In the high-SNR test the algorithm achieved
99:3 % sensitivity and 100:0 % PPV on average, whereas it scored 96:7 % sensitivity
and 91:3 % PPV in the low-SNR condition. Concerning the analysis performed
through the Ledalab software, a preliminary visual inspection analysis showed
a clear underestimation of both the tonic and the synthetic SMNA input signal
Fig. 3.3. All peaks, visually detectable after a zoom of the estimated phasic driver
component Fig. 3.4, results under the applied threshold. Despite all, by applying a
different threshold in the high-SNR test, the algorithm achieved 46:57 % sensitivity
and 67:03 % PPV on average, whereas no statistical consideration can be made about
the low-SNR condition. Furthermore, the sparsity of the phasic CDA signal is not
remarkable. While cvxEDA shows a robust behavior, the performances of Ledalab
software are extremely dependent from the additive noise power level.
4.0
simulated EDA, y
3.0 t
simulated
tonic
[a.u.]
2.0 simulated
SMNA
1.0 p
0.0
4.0
3.0
[a.u.]
2.0
1.0
0.0
0 10 20 30 40 50 60 70 80 90
τ [s]
Fig. 3.2 Solution of EDAcvx model applied to a synthetic signal with different levels of additive
white Gaussian noise (top: 33 dB; bottom: 13 dB)
4.0
3.0
[ a . u].
2.0
1.0
0.0
4.0
3.0
[ a . u].
2.0
1.0
0.0
0 10 20 30 40 50 60 70 80 90
τ [s]
Fig. 3.3 Phasic SMNA estimation using a CDA model applied to a synthetic signal with different
levels of additive white Gaussian noise (top: 33 dB; bottom: 13 dB)
0.3
0.2
[a.u.]
0.1
−0.1
0 10 20 30 40 50 60 70 80 90
τ [s]
0.3
0.2
[a.u.]
0.1
−0.1
0 10 20 30 40 50 60 70 80 90
τ [s]
Fig. 3.4 Solution of CDA model applied to a synthetic signal with different levels of additive
white Gaussian noise (top: 33 dB; bottom: 13 dB)
3.5.2 Experiment 1 Results
A visual inspection analysis of time series recorded during the forced maximal
expiration protocol confirmed the effectiveness of the paradigm in eliciting strong
SCRs that were partly overlapped because of short ISIs (Fig. 3.5). After applying
the cvxEDA algorithm, we considered peaks in the p signal within a 5-s time
window post-stimulus (considering the latency of a typical SCR [15]). Inter-subject
analysis indicated that the algorithm was able to identify the corresponding phasic
peak after 96:5 % of the stimuli and overcome the overlap issue. Furthermore, a
visual inspection analysis of the raw SC data in the time windows after stimuli
that were not identified by the algorithm showed the almost complete absence
of a SCR, probably because of incorrect performance of the task by the subject.
Results gathered from the CDA show a percentage of identification of the respiratory
stimulus of 93:01 %. The difference was due to some consecutive overlapped
stimuli, which were not recognized by CDA.
10
8
y [μS]
2
0
8
6
p [μS]
2
phasic driver [μS]
0
8
0
0 50 100 150 200 250 300 350 400
τ [s]
Fig. 3.5 Example of SC raw data (top) and its estimated phasic component by cvxEDA (middle)
and CDA (bottom) models during the forced maximal expiration task. Dotted lines mark the onset
of the visual cue triggering a forced expiration
3.5.3 Experiment 2 Results
Statistical analysis of the tonic and phasic driver components confirmed the
ability of the cvxEDA algorithm to characterize the ANS activity. Page-test results
comparing the four arousal sessions indicated a strong significant (p D 106 ,
L D 428, L D 4:74) relationship between the arousal level and the phasic driver
peak amplitude (see also Fig. 3.6(left)), which is the most appropriate parameter to
quantify ANS activity [15]. Post-hoc Bonferroni-corrected pair-wise comparisons
revealed significantly larger phasic responses to A2 than to A1 (p D 0:040,
Z D 2:47), to A3 than to A1 (p D 0:018, Z D 2:75), to A4 than to A1 (p D 0:002,
Z D 3:32), and to A4 than to A2 (p D 0:004, Z D 3:21). Concerning the tonic
component, the Wilcoxon test showed that the tonic mean values were significantly
higher during arousal than during neutral sessions (p D 0:001, Z D 3:147,
ZN D 0:813).
cvxEDA CDA
Mean ± SE Rank 4.0
3.0
2.0
1.0
A1 A2 A3 A4 A1 A2 A3 A4
* **
* *
**
Arousal Level
Fig. 3.6 Within-subject ranks of the peak amplitudes of the phasic component obtained by
cvxEDA (left) and CDA (right) for the four arousal levels. Dots mark the across-subject average
rank for each level while the wiskers indicate the standard error. The hypothesized effect of arousal
level on the phasic component was confirmed by the Page test (p D 106 for cvxEDA, p D 0:001
for CDA). Post-hoc Bonferroni-corrected pair-wise comparisons of the peak amplitudes found
significant differences in the cases indicated by asterisks ( W p < 0:05I W p < 0:01)
CDA estimated a phasic driver whose peak amplitude within each arousal
session increased with increasing levels of arousal (Page test p D 0:001, L D
409, L D 3:04). However, a post-hoc analysis revealed a significant difference
(p D 0:010, Z D 2:93) only between the most extreme sessions, A4 and A1
(see Fig. 3.6(right)). Overall, the new approach provides a stronger correlation
and augmented discriminant power, with respect to the elicited arousing session,
than CDA.
Chapter 4
Emotions and Mood States: Modeling,
Elicitation, and Recognition
In this chapter, we introduce basic concepts related to the theory of emotions, as

well as the strict link between emotions and mood/mental disorders. Then, ANS
correlates of emotions and mood disorders, with a special emphasis on EDA,
will also be reported. This knowledge backgrounds the experimental applications
described in details in the Chap. 5.
4.1 Theory of Emotions
Emotions are usually defined through features related to facial expressions, lan-
guage, gestures or posture, rather than their essential nature. This is due to the
enormous complexity of emotional phenomena that are often interrelated [94–98].
Moreover, they are the result of the continuous adaptation of individuals to changes
in the social environment. Indeed, the simultaneous coexistence of multiple positive
or negative emotions makes developing an accurate system of recognition of
emotions very complicated.
Nowadays, a remarkable theory of emotion belongs to the neurobiologist Joseph
LeDoux, who believed that the perception of an affective stimulus activates two
neural information paths that are independent but interconnected: an unconscious
path activating the peripheral reactions, and a conscious one producing the experi-
ence cognitive awareness of emotion [99]. According to LeDoux, one of the neural
centers related to the unconscious way is the amygdala. This region makes a first
process, of the stimulus perception, even with absence of awareness. Therefore, the
peripheral activation could be emotion-specific even before the complete conscious
experience. This mechanism is important in an evolutionary contest: considering
a potentially dangerous stimulus, the unconscious processing allows to implement
adaptive responses such as we have time to prepare the most appropriate response.
After a more detailed analysis of the stimulus that occurs through the sensory

46 4 Emotions and Mood States: Modeling, Elicitation, and Recognition
cortices and the conscious processing, in case the stimulus is not considered really
dangerous, the system falls rapidly to the basal state, otherwise the activation of
the emotional response is supported and retained. The presence of a direct visual
pathway that overcomes the visual cortex has recently been demonstrated in blind
patients due to a partial or total destruction of the primary visual cortex. Although
they could not consciously see anything, they had a significant activation of the
amygdala in MRI, when they presented images of affective face expressions [100].
In LeDoux’s theory, the conscience of the stimulus does not result from the periph-
eral response, and the awareness of emotion does not cause a peripheral change.
Rather, the perception of an affective stimulus determines both the peripheral
change (through a subcortical region that is inaccessible to consciousness) and the
awareness of emotion. The unconscious and conscious systems affect each other
increasing or decreasing their activity.
What are emotions? A definition that partly reflects LeDoux’s theory describes
them as a quick, automatic, and stereotyped response of the organism to a
potentially significant stimulus to the survival of the individual or the species.
Certain situations can be unpleasant or dangerous to the survival of every individual.
For this reason it is necessary for the body to react as quickly and effectively as
possible. For example, in a very dangerous situation we do not often have time to
think, or rather the body must be prepared to react before we start thinking. To
make this possible, our physiological response must be automatic and consequently
stereotyped, identical or very similar in different situations in which we experience
the same emotion: fear of a bear can not be different from the fear of a lion or a
robber. The last aspect is probably related to the fact that emotions (at least those
considered primary ones as anger, surprise, happiness, sadness, fear and disgust) are
not only present in humans but are universally expressed in the same way. Studies
on facial expressions by the American psychologist Paul Ekman confirm this idea:
we can feel fear for different things, but always in the same way [101]. Of note, the
time profile of the emotions changes rapidly both in the onset, and during the phase
of extinction. In fact, when the emotive situation is concluded, typically the emotion
ends in both the cognitive and peripheral psycho-physiological component.
4.2 Modeling Emotions
In the literature, several theories for modeling emotions have been proposed.
Discrete, dimensional, appraisal and dynamical models are the most interesting, but
one cannot exclude the others.
In discrete models, emotions can be seen as the result of a selective adaptation
that ensures survival [102]. This survival concept could be illustrated by the
following relation: danger D> fear D> escape D> survival. The result of this
selection is a small set of basic, innate and universal emotions. For instance, Ekman
proposed 6 basic emotions which are identified on the basis of facial expressions:
anger, disgust, fear, joy, sadness and surprise [103, 104]. Besides, in the literature
other discrete models have been proposed and they include more or less basic
4.2 Modeling Emotions 47
emotions, usually from 2 to 10 [105–107]. These emotions are called primary

emotions as opposed to secondary emotions which result from a combination of
the primary ones (e.g., contempt D anger C disgust). Nevertheless, this model can
be insufficient to describe mixed emotions which necessarily require much more
than one word to be expressed. Moreover, there are controversies in the assumption
of universality of basic emotions (Darwinian hypothesis [102]). What seems true is
that emotions are universally expressed (e.g. facial expressions [108]) but dependent
on semantic attributions. It is demonstrated that inter-cultural differences, e.g.
difference between asian and occidental people, are more important than intra-
cultural differences, e.g. between genders, and that no significant differences
between primary and secondary emotions exist. From an evolutionary point of view,
basic emotions may be the first emotions infants could experience [109] (see Ortony
et al. [110], for basic emotion categories defined over the years).
Unlike discrete models, dimensional models consider a continuous multidimen-
sional space where each dimension stands for a fundamental property common to all
emotions. This kind of model has already been used by Wilhelm Wundt [111]. Over
the years, a large number of dimensions has been proposed [112–117]. Two of the
most accepted dimensions were described by Russel [118]: valence (i.e. pleasure,
positive versus negative affect), and arousal (low versus high level of activation).
These dimensions derived from a valence, arousal, and dominance space developed
by Russell and Mehrabian [119], in which dominance represents the degree of
control over the situation.
Appraisal models are based on the evaluation of current, remembered or imag-
ined circumstances. At the heart of the appraisal theory is the idea that the particular
judgements made about the environment and ourselves cause different emotions.
The situational appraisals appear to be highly dependent on motives and goals.
In other words, how we feel depends on what is important to us, indeed all our
appraisals are connected to what we want and, therefore, to how we feel. For
example, frustration results from a goal which is not achieved. This model was
introduced by Arnold [120] and has been developed and refined by Frijda [94],
Ortony et al. by creating the OCC model [121], Scherer with the Component
Process Theory [122] and the derived one by Lisetti and Gmytrasiewicz [123]. The
appraisal process can be thought as having a continuous and a categorical nature.
Roseman’s model (1996) shows that appraisal information can vary continuously
but categorical boundaries determine which emotion will occur. To solve the
problem between categorical and continuous appraisal order, it may be a good
idea to place discrete emotional categories (i.e. happiness, sadness, etc.) while
continuous models represent varieties, styles, and levels of these already defined
distinct emotions [124].
Finally, the dynamical model approach considers emotions as a dynamical
process. This model starts from an evolutionary perspective and characterizes
emotion in terms of response tendencies. In the dynamics, a perspective emotion is a
regulable system and the capability of understanding its rules is essential. According
to a process model of emotion regulation, emotion may be regulated at five points
in the generative process: selection of the situation, modification of the situation,
deployment of attention, change of cognitions, and modulation of responses. It may

be useful to take into account concepts like mood and personality (see Egges et al.
[125]).
In all of the studies presented in this book, a common dimensional model com-
prising multiple dimensions to categorize emotions is employed: the Circumplex
Model of Affects (CMA) [126]. This model interprets the emotional mechanisms
underlying affect as a continuum state. They are represented on a Cartesian system
of axes, each of which refers to a neurophysiological pathway of emotional
processing. In many cases, using factor analysis and multidimensional scaling of
a wide set of psychometric assessments and self-reports on emotional states, it is
possible to employ a more simplified bi-dimensional model. In particular, the two
dimensions can be conceptualized through the terms of valence and arousal, which
are usually intended as the two independent, predominantly subcortical systems that
underlie emotions (see Fig. 4.1). Valence represents how much an emotion is felt
by people as positive or negative. For example, someone feeling sad has evaluated
surrounding events as very negative. On the contrary, someone feeling joyful would
have appraised the environment as positive for his well being. Arousal indicates how
relevant the surrounding events are and therefore how strong emotion is. In this case,
someone feeling excited will have an emotion represented by a bigger arousal and
someone feeling bored will experience a much less relevant emotion. Accordingly,
in CMA, arousal and valence can be considered adequate dimensions to identify
specific emotions.
4.3 Autonomic Nervous System Correlates of Emotions
It has been demonstrated that ANS dynamics reflects measurable changes according
to subjects’ emotional experience [126, 127].
ANS is a control system in charge of the regulation of peripheral functions
such as heart rate, digestion, respiratory rate, pupillary response, urination, and
sexual arousal [128]. This system comprises two components: the sympathetic
and parasympathetic nervous systems. Due to its aspecific nature, the ANS is not
involved only in emotion regulation, but includes a wide variety of other functions
related to stress, attention and so on [129]. As a matter of fact, several physiological
ANS signs (e.g. HRV, respiration activity, EDA, pupil size and eye movement
variation) correlate with subject behavior or emotional status [130–141]. The most
commonly used indexes of activation of the ANS are based on EDA (i.e., sweat
glands) or cardiovascular dynamics [142]. As mentioned in the previous chapters,
EDA is typically quantified in terms of skin conductance and primarily reflects
sympathetic activity.
In 1884, James was the first psychologists to claim that different emotional
states (such as sadness, anger or fear) involve specific parameters of activation
of the autonomic nervous system [143]. This principle has been very important
in many theories of emotions [144, 145], and most of the researches inspired by
4.3 Autonomic Nervous System Correlates of Emotions 49
Arousal
Alert
Tense
ACTIVATION
s Ex
ou ict
ed
N erv
ed
El
ss
ate
re
d
St
Happ
Upset
y
UNPLEASANT PLEASANT
Valence
nted
Conte
Bor
DEACTIVATION
ed
e
D
ren
ep
re Se
ss
ed d
la xe
Sad Re
Calm
Fig. 4.1 A graphical representation of the circumplex model of affect with the horizontal axis
representing the valence dimension and the vertical axis representing the arousal or activation
dimension
James’s theory have focused on ANS measures. The strong scientific interest in the
specificity of the ANS is due to popular thought that emotions involve discrete types
of activation of the ANS (e.g., the supposed link between anxiety and increased
heart rate [146]). However, there are conflicting positions in the scientific society.
Although some works have reported evidences for specificity of the ANS [147–149],
a recent meta-analysis has featured such effects as inconsistent [150]. In this meta-
analysis, only a small part of the analyzed ANS correlates reliably distinguished
discrete emotions, highlighting the lack of ability to support the hypothesis of
specificity of ANS [150].
Given these considerations, Lang et al. [90] have shown in some studies that
the level of skin conductance increases systematically and linearly depending on
the general arousal level of some emotional stimuli. In addition, the relationship
between affective stimuli and EDA is note independent from the valence, from the
kind of stimulation, and certainly from the specific emotion that was originated.
These findings were consistent with theories that argue that ANS activities indicate
the level of excitement of the emotional state rather than its basic emotion
[151, 152]. However, not all measurements of the ANS can be mapped in a single
dimension. In accordance with the principle of “directional fractionation” [153],
different measurements of ANS activity can operate independently or contrariwise.
For example, the heart rate decreasing can contribute to an increase in sympathetic
activity as assessed by other ANS correlates [154]. In order to explain such
fractionation of the ANS, at least two dimensions should be considered, e.g., the
valence. For example, Cacioppo et al. [150] revealed that cardiac output, blood
pressure, heart rate and skin conductance respond to the emotional valence.
Although the ANS seems sensitive to dimensional rather than to discrete emo-
tional states, taking in account many ANS correlates can help achieving a greater
degree of autonomic specificity [155]. For example, anger and fear, despite their
alignment in terms of valence and arousal, can be differentiated by a combination
of cardiovascular and respiratory measures [142]. Thus, combinations of several
measures ANS can provide a better (automatic) detection of discrete emotional
states.
4.4 Affective Computing
The automatic quantification and recognition of human emotions is a relatively new

and fast-growing research area which combines knowledge in the fields of psycho-
physiology, computer science, biomedical engineering, and artificial intelligence.
Results of these studies are usually identified within the so-called “Affective Com-
puting” field, providing computational models and machine learning algorithms for
the automatic recognition of emotional regulation occurring through different kinds
of elicitation. In general, an emotion recognition system is designed to be effective
for a specific kind of stimulus and it is built on a specific model of emotion which
has to be characterized by processing one or more physiological/behavioral signs.
Recently, several engineering approaches have been used in order to guarantee
that acceptable emotion recognition systems have high accuracy, robustness, and
adaptability to practical applications. An emotion recognition system generally
comprises two main parts: emotion elicitation and physiological correlates iden-
tification.
Such systems are devised to map physiological patterns into well-defined emo-
tional states for an automatic classification. The physiological signs include implicit
and explicit emotional channels of human communication, such as speech, facial
expression, gesture, physiological responses [156]. Recently, numerous automatic
emotion recognition systems have been proposed involving, among others, human
or patient-robot interactions [157, 158], car drivers [159, 160], facial expression
[161], and adaptation of game difficulty [162]. Table 4.1 summarizes the most
relevant results reported in literature during the last decade about the emotion
recognition through the ANS biosignal response [131–133, 159, 163–175]. All the
acronyms used in this table are expanded in the Acronym-Table at the beginning of
this book. Each row of the Table 4.1 shows the first author along with the publication
year, the set of physiological signals used for that study, the typology of stimulation
pattern, the emotion classes, the type of the classifier and the results in terms of best
percentage of successful recognition.
4.4 Affective Computing 51
Table 4.1 Performance of the peripheral biosignal based emotion recognition methods reported in
the literature of last decade
Authors Signals Elicitation Emotion classes Accuracy Best results (%)
Yoo et al. ECG, EDA Video clip Sad, calm ANN 80
[166] pleasure,
Interesting
pleasure, Fear
Choi & Woo BVP, EDA Music and Joy, anger, and ANN 74.5
[167] image sadness
choosen by
subject
Healey & EMG, ECG, Driving 3 Stress levels LDA 97
Picard [168] EDA, RSP
Li & Chen ECG, BVP, Film clips Fear, neutral, and CCA 93.33
[169] EDA, ST joy
Rani et al. ECG, BVP, Cognitive Engagement, SVM 86
[170] EDA, EMG tasks (i.e. anxiety,
Anagrams boredom,
and Pong) frustration and
anger
Rainville ECG, RSP, Self induction Anger, fear, SDA 49
et al. [171] EDA, EMG happiness,
sadness
Zhai & EDA, BVP, Stroop test 2 stress levels SVM 90
Barreto [172] PD, ST game
Leon et al. ECG, EDA, IAPS Neutral, negative, ANN 71
[173] BVP positive
Liu et al. ECG, ICG, Cognitive Anxiety, SVM 83
[174] BVP, HS, tasks (i.e. engagement,
EDA, EMG, anagrams and liking.
ST Pong.)
Katsis et al. EMG, ECG, Car-racing High stress, low SVM 79.3
[159] RSP, EDA drivers stress,
disappointment,
euphoria
Yannakakis & ECG, BVP, Interactive 2 fun levels SVM, 70
Hallam [175] EDA games ANN
Kim & André EMG, ECG, Music 4 musical LDA 70/95
[132] EDA, RSP listening emotion
Katsis et al. BVP, ECG, IAPS Relaxed, neutral, ANN, 84
[133] EDA, RSP startled, SVM
apprehensive,
very
apprehensive
Gouizi et al. EMG, RV, IAPS 6 discrete SVM 75–83
[176] SKT, EDA, emotion
BVP, ECG
(continued)
Table 4.1 (continued)

Authors Signals Elicitation Emotion classes Accuracy Best results (%)
Valenza et al. ECG, EDA, IAPS Arousal, QDC 90/92
[177] RSP valence
Jang et al. BVP, EDA, Video clips 4 negative LDA, 51–100
[178] ECG emotions CART,
NN,
SVM
Lanata et al. EDA IAPS 5 Arousal levels QDC 94
[57]
Ren et al. EDA Mental stress “Relaxation” SVM 60.66
[179] task vs. “stress”
states
Torres et al. EMG, RESP, Music videos Arousal, SVM 70/60
[180] SKT, EDA, valence
BVP, EEG
Yanga et al. EDA Emotional 4 discrete Fisher, 69–80
[181] movies emotions KNN,
LDC,
QDC
Maaoui et al. EMG, RESP, IAPS Valence SVM 91
[182] SKT, EDA,
BVP
Kukolja et al. ECG, SKT, IAPS 5 discrete MLP 60
[183] EDA emotions
Harinaran EDA, ECG Controlled Valence CART 67
et al. [184] trading
experiment
Khezri et al. BVP, EDA, Video clips 6 basic SVM, 85/80
[185] EEG, EMG Ekman’s KNN
emotions
Khan & Lawo BVP, EDA IAPS 8 discrete Decision 92
[186] emotional tree
states
Rukavina EMG, ECG, IAPS 3/ 5 SVM 54
et al. [187] EDA arousal/valence
classes
4.5 Multi-Sensory Elicitation
How emotions can be elicited to humans is a crucial issue still open. Many
researches have turned their attention to sensory stimuli able to elicit emotions.
The difficulty associated with the elicitation is related to a complex interaction
between cognition and neurophysiological changes. Several modalities and percep-
tual channels could be used for this purpose, which can be thought as affected by
several “noisy” factors, including psychophysiological processes such as attention,
social interaction, and body-to-biosensors connections. In the literature, a wide
4.6 Emotions and Mood Disorders: Bipolar Disorder 53
range of elicitation methods have been applied: introspection, movements, lights and
colors [188], set of actions, images (e.g IAPS described below) [90, 189], sounds
(e.g., music and IADS described below) [132, 190–193], (fragments of) movies
[194, 195], speech [196], commercials [197], games, agents/serious gaming/virtual
reality [173], reliving of emotions [198], real world experiences [168, 199] along
with using personalized imagery stimuli [131].
In order to evoke affective states in a laboratory setting, some authors have
assembled sets of pictures [90], sounds [200], odorants [201], and words chosen
to elicit a range of positive, neutral or negative affective states. In such studies,
subjects had to rate pictures, sounds, odorants or words in terms of pleasure and
arousal. Results indicated that the shape of the distribution is very similar across all
sensory stimulations [142].
In this perspective, the International Affective Picture System (IAPS) [90] and
the International Affective Digital Sounds system (IADS) [202] are two of the most
frequently used tools in the area of affective elicitation. They consist of hundreds of
images and sounds with associated affective scores. A commonly used approach is
to have a collection of stimuli where each one is slightly varied in terms of intra-
individual standard deviation of affective ratings.
In several experiments reported in this book, a set of images gathered from the
IAPS was chosen [203]. IAPS is a set of 944 images having a specific emotional
rating, in terms of valence, arousal, and dominance. The emotional ratings are
based on several studies previously conducted where subjects were requested to
rank these images using the self assessment manikin [204]. The elicitation by
IAPS is able to activate segregated neural representations of the different emotion
dimensions in different prefrontal cortical regions [205, 206]. Touch is another
sense capable of inducing emotions. Specifically, affective haptics is the science
referring to the ability of haptic systems to communicate emotions, possibly
affecting social behavior and interactions [207]. This is possible because of the
action of a specialized kind of tactile sensors in the skin, i.e., the unmyelinated
CT tactile fibers [208, 209], whose activity is linked to the controlateral primary and
bilateral secondary somatosensory area, as well as contralateral middle and posterior
insula cortex [210]. Previous studies demonstrated how these fibers are sensitive to
changes in the physical characteristics of the haptic stimulus. Specifically, changes
in contact force and velocity of human caresses can vary the valence perception
(pleasantness/unpleasantness) of the stimulus [211, 212].
4.6 Emotions and Mood Disorders: Bipolar Disorder
Bipolar disorder is a chronic illness involving millions of people in Europe and in

the United States (see the epidemiological study in [213]). Patients experience mood
swings whose symptoms can be associated to one of the following psychophysiolog-
ical states: depressive, maniac, mixed, and euthymic. During depressive episodes,
patients feel sad and, sometimes, desperate. Other neurovegetative symptoms
including loss of appetite and sleep are also present. Depressed patients might also
experience thoughts of ruin, guilt or death including suicidal thoughts that might
lead to suicide attempts. During manic episodes, patients are hyperactive, and often
experience a reduction of the need to sleep. Mixed states are characterized by both
depressive and hyperactivity symptoms. In the intervals between these episodes,
patients typically experience periods of relatively good emotional balance (labeled
as euthymia). Moreover, mood swings are also usually accompanied by anxiety,
which is associated with bipolar disorder either as a symptom of the bipolar disorder
itself or as a separate pathological condition [214].
Despite the great impact of bipolar disorder on the population and healthcare
costs, current clinical practice still relies only on the physician expertise, rating
scales and questionnaires, such as the Bauer Internal Mood Scale, the Hamilton
Scale for Depression and the Young Mania scale [215]. Physiological parameters
(e.g., biological markers, physiological signals, etc.) are not taken into account
for diagnosis or follow-up purposes [216–218]. As a matter of fact, there is the
need of more objective parameters for the diagnosis of mental disorders. These are
long-term illnesses and may remain undetected for years before they are properly
diagnosed and put under treatment. Moreover, patients are extremely heterogeneous
with respect to the phenomenology and severity of symptoms, number and duration
of episodes, as well as time interval between them. Finally, other disorders may also
be present (i.e., comorbidity).
Previous researches have shown a link between Autonomic Nervous System
(ANS) dysfunctions and bipolar disorders [8, 219–222]. Specifically, studies on
sleep [223], voice analysis [10], and circadian heart rate rhythms [224, 225] showed
to be sensitive to changes in clinical state, suggesting that these parameters may be
considered as markers of clinical change. Moreover, it is known that electrodermal
hypoactivity is present during depression in both unipolar and bipolar patients
[226, 227]. This condition is stable over time, and does not appear to depend on
experimental conditions or stimulus characteristics [228].
Since changes on EDA are directly related to the sympathetic activity [23], EDA
analysis could serve as an effective ANS marker for characterizing different mood
states.
The study on bipolar patients presented in this book was carried out in the
frame of the European project PSYCHE, which stands for personalized monitoring
systems for care in mental health. Within such a project, a personalized, pervasive,
cost-effective, and multi-parametric monitoring system based on textile platforms
and portable sensing devices was devised for the long-term and short-term analysis
of mood disorders [8–10].
Chapter 5
Experimental Applications on Multi-Sensory
Affective Stimulation
In this chapter, we report in detail several experimental methods concerning affec-

tive stimulation and results gathered applying EDA models to SC data. In particular,
we focus on two specific research fields: emotion recognition and assessment of
mood/mental disorder.
Emotion Scenario The scientific debate on the physiological origin of emotions
is still open: whether they originate from the peripheral reactivity of the ANS, or
from specific areas of the brain, or from both. As Damasio stated, “emotions are
the most complex expression of homeostatic regulatory systems”. He hypothesized
that emotions (or emotional memories) can modify our behavior through conscious
or unconscious signals [229]. Note that the latter ones belong to the ANS signaling
whose role is to generate reentry vegetative information to pre-existing cortical maps
[229, 230].
Human emotions involve several areas of the cerebral cortex for their regulation
and feeling. The prefrontal cortex and amygdala, in fact, represent the essence of
two specific pathways: affective elicitations longer than 6 s allow the prefrontal
cortex to encode the stimulus information and transmit it to other areas of the
Central Autonomic Network (CAN), thus producing a context appropriate response
[231]; briefly presented stimuli access the fast route of emotion recognition via
the amygdala. Regulatory areas of the CAN include anterior and midcingulate
cortices, insula, ventromedial prefrontal cortex, mediodorsal thalamus, amygdala,
and hypothalamus [232].
Dysfunctions on these CNS recruitment circuits lead to pathological effects [233]
such as anhedonia, i.e. the loss of pleasure or interest in previously rewarding stim-
uli, which is a core feature of major depression and other serious mood disorders.
Given the CAN involvement in emotional responses, an important direction for
affective studies is related to changes of the Autonomic Nervous System (ANS)
activity as elicited by specific emotional states.
The automatic quantification and recognition of human emotions is a relatively
new and fast-growing research area which combines knowledge in the fields

56 5 Experimental Applications on Multi-Sensory Affective Stimulation
of psycho-physiology, computer science, biomedical engineering, and artificial

intelligence. Results of these studies are usually referred to as the so-called
“Affective Computing” field, providing computational models and machine learning
algorithms for the automatic recognition of emotional regulation occurring through
different kinds of elicitation [234]. Concerning the affective elicitation, a wide range
of elicitation methods have been proposed in the literature: real experiences [199],
film clips [5, 169, 235, 236], problem solving [237], computer game interfaces [238],
images [165], as spoken words [239] and music [240–242].
Given the physiological reasons highlighted above, a wide class of affective
computing studies focused on investigating how the ANS activity changes upon
specific emotional stimuli. A recent review written by Calvo et al. [156] reports
on emotion theories as well as on affect detection systems using physiological and
speech signals (also reviewed in [243]), face expression and movement analysis.
In the field of affective computing, the Electrodermal Activity (EDA) has been
extensively proposed in the literature along with other ANS-derived signals such
as heart rate variability (HRV), diastolic and systolic blood pressure, pupillary
dilatation, respiration, temperature and skin conductance [8, 15, 127, 165, 193, 222,
236, 244–248].
EDA has been closely linked to autonomic emotional and cognitive processing,
and it is widely used as a sensitive index of emotional processing and sympathetic
activity. Among the ANS correlates, EDA has been one of the favorite tools in
this context. Partly, this can be attributed to the ease of its measurement. Only a
simple electronic circuitry is required to record it. One of the major reasons for its
importance lies in the fact that EDA is solely mediated by the sympathetic branch
of the ANS, thus being not subjected to peripheral parasympathetic influences as
most of the other autonomic measures. Although most of the affective computing
studies which can be found in the literature are based on a multiparametric approach,
i.e. several signals from ANS dynamics, in this book we propose an automatic and
efficient emotion recognition system based exclusively on measures derived from
EDA.
Mental Health Scenario We applied the EDA analysis also to data gathered
from patients with bipolar disorders [8–10]. These patients experience a series of
mood swings among depressive and manic or hypomanic episodes throughout their
life. Although common and expensive to treat, the clinical assessment of bipolar
disorder is still ill-defined. Since the current literature reports on several correlations
between mood disorders and dysfunctions involving the ANS, our objective was
to develop a novel and reliable mood recognition system based on EDA analysis.
The ANS monitoring platform used for this study is the core sensing of the
personalized monitoring systems for care in mental health (PSYCHE) European
project [8, 11]. Bipolar patients were followed for a period of 90 days during which
up to five monitoring sessions and psychophysical evaluations were performed for
each patient. Specific experimental paradigm using visual emotional stimuli was
performed. Statistical analysis of the phasic features corresponding to each mood
state demonstrated how EDA analysis is able to provide effective biomarkers for the
5.1 Multi-Sensory Experimental Applications 57
assessment of pathological mood states in bipolar disorder. EDA features were also
inputed to an automatic classifier for mood state recognition, achieving an accuracy
> 80 % for the inter-subject analysis while discerning between the state of good
affective balance (euthymia) and severe mental states such as depression and mixed
state. Therefore, experimental evidences on the correlation between pathological
mood disorders and EDA were found, and the obtained results are promising for an
effective and objective biosignal-based mood recognition [9, 11].
5.1 Multi-Sensory Experimental Applications
In the next sections we report a set of experimental applications in the field of

affective computing and emotion recognition. The experiments are intended to
induce emotional stimuli and record the correspondent physiological response in
both healthy subjects and bipolar patients. In addition, results of an experimental
paradigm designed in order to test differences between DC and AC measurement
approaches are discussed.
The rationale behind all of the proposed procedures is to maximize the ANS
dynamical response under specific stimuli with respect to the baseline (rest condi-
tion). The affective elicitation was performed by means of multi-sensory stimuli.
Specifically, we studied the ANS response to visual, auditory, tactile, and olfactory
affective stimulations through the EDA investigation, applying both the cvxEDA
and CDA models.
It is well known that a physiological response, characterized by different levels
of arousal and valence (according to the CMA model), is inducible using visual
and acoustic stimuli, e.g., IAPS images and IADS sounds [249]. Several studies
demonstrated also that haptic and olfactory stimuli can produce an affective
response [208, 250]. One crucial aspect in the affective computing field consists
in the standardization and reproducibility of the stimuli. The use of standardized
stimuli allows replicating studies in a more reliable fashion, making also easier
the comparison of the results with future studies. Therefore, each stimulation was
characterized in terms of valence and arousal following the Russell’s CMA model
described in Chap. 4. Upon these considerations, the visual and acoustic sets were
comprised of pictures and sounds extracted form the IAPS and IADS database,
respectively. Both IAPS and IADS stimuli are characterized in terms of valence
and arousal dimensions.
Concerning the tactile and olfactory stimulations, to our knowledge, in the
literature there are not standardized databases. Therefore, in this study we utilized
standardized stimuli using a haptic device able to simulate the human caress
with different force and velocity levels, and five smells synthesized in laboratory
according to some guidelines from the literature. After each stimulus, each subject
was asked to rank the tactile and olfactory stimulus in terms of valence and arousal.
5.2 Classification Procedure
Statistical analysis and pattern recognition methodologies were used to automati-

cally recognize the autonomic response of the subjects to the emotional stimulus,
and to associate the values of the EDA parameters with the kind of elicitation
as expressed in terms of arousal and valence. This part of the study has been
implemented following a Leave-One-Subject-Out procedure (LOSO): we applied
a feature selection to each training set comprised of .N 1/ subjects (where N is
the total number of participants) to recognize the emotional responses of the subject
Nth. This procedure was iterated N times.
The statistical analysis was intended as a preliminary feature analysis procedure
aimed at identifying the parameters that significantly vary among the arousal
and valence levels. Before performing the statistical analysis, Shapiro-Wilk test
was performed in order to check whether the data were normally distributed. In
case of non-normal distribution, the results are expressed in terms of median and
Median Absolute Deviation (MAD). Consequently, the Friedman test [251], i.e.
nonparametric one-way analysis of variance for paired data, was used to test the
null hypothesis that no difference exists among different sessions (more than two),
as well as the Wilcoxon signed-rank test, i.e a non-parametric test used to compare
paired samples from two sessions, to assess whether their population medians
differ. We applied these statistical tests in order to discern the arousal sessions,
and the negative and positive valence levels in each experimental protocol. In the
next sections, for each statistical analysis related to a specific sensory stimulation
protocol, we will specify in detail the features extracted and the kind of statistical
comparisons performed.
Concerning the emotion recognition problem, two different classification
approaches were implemented according to the nature of data, i.e., paired or
unpaired.
5.2.1 Paired Within-Rank K-NN Classifier
In all of the experimental applications described in this book, each participant took
part to all the different arousal and valence level sessions. Consequently, to compare
the arousal and valence stimuli, a paired dataset has to be considered. Therefore,
a non linear classifier for paired data was implemented and used. This classifier,
called paired within-rank (PWR) K-NN, acts in two steps: first, transforms the
features set in a within-subject rank matrix based on the Friedman test statistic; then,
the ranked matrix is applied as input of a k-Nearest Neighbor algorithm (K-NN)
classifier.
In detail, the dimension of the feature space was given by the number of selected
EDA parameters. The final features set was achieved identifying the group of
parameters that performed a higher recognition accuracy in the training set. For the
5.2 Classification Procedure 59
Decomposition process
Phasic cvxEDA
and or EDA
Tonic CDA
Feature Extraction
and Statistical Analysis
Leave One Subject Out

Procedure
Training Set
Feature
of N-1
Selection
subject
Test Set: Pattern

i-th subject Recognition
Fig. 5.1 Overall block scheme of the proposed emotion recognition system. The EDA is processed
in order to extract the phasic and tonic components. According to the protocol timeline, several
features are extracted and, then, statistically compared. The PWR-KNN algorithm is engaged to
perform pattern recognition by adopting a leave-one-subject-out procedure
arousal and valence recognition, the number of samples (i.e., rows of the dataset)
in such a space was related to the number of subjects multiplied by the number
of arousal or valence levels to be classified. Before the classification procedure,
the values of each feature (i.e. a column of the dataset) were transformed ranking
the data within each subject. Finally, the ranked feature set was used as input to
a K-NN Classifier [252, 253], which was validated through the LOSO procedure
[254], as described at the beginning of this section. A block diagram of the proposed
recognition system is illustrated in Fig. 5.1.
The K-NN is considered one of the simplest machine learning procedures. An
example is classified considering the majority vote of its neighbors K, which
is a positive integer typically not very large. The choice of K depends on the
characteristics of the data. Generally the increase of K reduces the noise that affects
the classification, but the criterion of choice for the class becomes more labile.
The choice is made by means of heuristic techniques such as the cross-validation.
The space is partitioned into regions based on the positions and characteristics of the
training examples. For the calculation of the distance, examples are represented by
vectors in a multidimensional space and the Euclidean distance is used. A point

(which represents an example) is assigned to the most frequent class among the K
nearest training-examples to the test one under examination.
K-NN is a non-parametric classification method; therefore, no assumption of the
underlying probability density functions is needed. It is assumed that there are ade-
quate data points in each class so that in any small area within the decision region,
the number of occurred data points in these areas imply the true characteristics of
each density function [255]. In comparison with more sophisticated classifiers, KNN
classifiers showed remarkably small error rates.
5.2.2 Support Vector Machine
The SVM method was used as an alternative method for recognizing the valence
and arousal levels in case of unpaired dataset (as in the tactile paradigm,
see Sect. 5.5.5.4). It replaces the ranked transformation function and the K-NN
method.
The classification problem can be viewed as a task of finding a separating
hyperplane that divides the examples belonging to each different classes. Also in this
case, the classification procedure consists of two phases: training and testing. Firstly,
the dataset is divided into two groups: the training set and the test set. During the
training, the algorithm estimates a hyperplane that separates the examples contained
in the training set according to their labels. The decision function that has been
learned from the training data then can be used during the testing phase to predict
the class of a new test example. More details about SVM can be found in [255].
Of note, all classification results are expressed in this book as recognition
accuracy in form of confusion matrices [256]. A generic element cij of a confusion
matrix indicates the percentage of how many times the feature set belonging to the
class i, was recognized as belonging to the class j. This means that a higher average
of the values on the matrix diagonal corresponds to a better degree of classification.
5.3 Affective Visual Elicitation
The most widely used affective stimulation is the presentation of images taken
form the IAPS database. They consist of hundreds of images, with associated
standardized affective values (in terms of arousal and valence). In the literature
there are several studies where IAPS database is used to estimate the affective
state (especially when the arousal level changes) through the study of physiological
signals, such as electroencephalogram [257–260], blood oxygen level dependent
signals [261], facial electromyograms [92, 262, 263], EDA [264–266] and HRV
[248, 260, 267]. Concerning EDA, many studies over the past years have demon-
strated that the amplitude of the skin conductance changes and the intensity of
emotional experience is almost linearly associated with the arousal dimension,
[92, 268, 269].
5.3 Affective Visual Elicitation 61
In the previous chapter (see Sect. 3.3), the relationships between the arousal
levels of the IAPS and the phasic component features coming from CDA and
cvxEDA models, have been already statistically compared. CvxEDA has shown a
statistical discriminant power higher than CDA. Indeed, all the four arousal levels
were significantly different and showed a trend with a high correlation with the
arousal scale. CDA significantly discriminated only the extreme levels but not the
middle ones, see Fig. 3.6.
Furthermore, a multivariate pattern analysis using the PWR KNN classifier was
performed in order to automatically recognize the four different levels of arousal
and two levels of valence.
5.3.1 Experimental Protocol of Affective Visual Elicitation
A group of 15 healthy subjects was recruited to participate in the experiment.

Their age ranged from 21 to 24 and were naive to the purpose of the experiment.
The affective elicitation was performed by projecting the set of IAPS images to
a PC monitor. The slideshow was projected in a room equipped with a dedicated
monitor and headset to acoustically insulate from external noise. The slideshow was
comprised of 9 sessions of images N, A1,N, A2, N, A3, N, A4, N (Fig. 5.2), where
N is a session of 6 neutral images (mean valence rating 6:49 , SD D 0:87 , range
D5:52 7:08 ; mean arousal rating D 2:81, SD D 0:24 , range D 2:42 3:22 ) , and
Ai (with i going from 1 to 4 ) are sets of 20 images eliciting an increasing level of
arousal and valence. Detailed values are reported in Table 3.1. The overall protocol
utilized 110 images. Each image was presented for 10 s.
Valence
Low Medium-low Medium-high High

Arousal Arousal Arousal Arousal
Neutral
Fig. 5.2 Timeline of the experimental protocol in terms of arousal and valence levels. The vertical
axis relates to the valence score, whereas the horizontal axis relates to the time
Table 5.1 Confusion matrix AR1 AR2 AR3 AR4

of arousal levels using the
CDA features set AR1 33.33 40.00 33.33 13.33
AR2 26.67 46.67 26.67 20.00
AR3 33.33 6.67 33.33 33.33
AR4 6.67 6.67 6.67 33.33
Results are presented as a percentage
value
The bold values represent the percentage
of number of predictions/classifications
that were correct (true positives and true
negatives)
Table 5.2 Confusion matrix AR1 AR2 AR3 AR4

cvxEDA features set AR1 73.33 13.33 6.67 6.67
AR2 20.00 66.67 6.67 0.00
AR3 6.67 20.00 66.67 13.33
AR4 0.00 0.00 20.00 80.00
value
The bold values represent the percentage
negatives)
Table 5.3 Confusion matrix Positive Negative

of valence levels using the
CDA features set Positive 68.33 31.67
Negative 31.67 68.33
Results are presented as a percent-
age value
The bold values represent the per-
centage of number of predictions/
classifications that were correct
(true positives and true negatives)
5.3.2 Classification of Visual Arousal and Valence Levels
Concerning the arousal levels, the multivariate pattern recognition analysis showed a
strong difference between the accuracy of the cvxEDA feature (71:67 %) set and the
CDA feature set (36:66 %) (see Tables 5.1 and 5.2). Instead, concerning the valence
levels, the two models demonstrated the same performance, discerning pleasant and
unpleasant images with a similar accuracy of 68:33 % (see Tables 5.3 and 5.4).
5.4 Affective Sound Elicitation 63

cvxEDA features set Positive 68.33 31.67
Negative 31.67 68.33
age value
5.4 Affective Sound Elicitation
One of the computational study proposed in this book aims to recognize emotions
through the identification of the elicited arousal and valence levels of standardized
acoustic stimuli gathered from the International Affective Digitized Sound system
(IADS) [270]. Likewise the IAPS database [202, 271], IADS is a database of
affective sounds characterized in terms of valence and arousal dimensions [270].
The auditory stimulus is one of the most powerful means to induce and
communicate emotions to people. It is easy to understand the role that voice tone
plays in conveying speaker affect [272], or in eliciting hearer emotions. The music
is another way to enhance orally expressed affective messages [273, 274]. Speech
and music seem to be only a portion of the sounds that we hear. There are also
non-musical and nonlinguistic sounds [275], which carry affective information in
the audio environment around a listener [276].
In the literature there are studies dealing with the relationship among physi-
ological signals and pleasant and unpleasant sounds or music-induced emotions.
PET and fMRI studies on emotions evoked by auditory stimuli have found that
pleasant sounds lead the activation of brain areas such as the orbitofrontal cortex
and the anterior insula [277]. Sad music or unpleasant noises instead lead to the
activation of regions involved in negative emotional states and anxiety-related,
such as the hippocampus, the amygdala and the areas of the medial temporal
lobe [278, 279]. All this brain areas are directly involved in the control of the
human affective system. Other studies take into account EEG and ANS dynamics
[192, 240–242, 280, 281], also to automatically recognize four types of music-
induced emotions [192]. In general, it is well known that there are specific peripheral
activation patterns associated with the emotional valence of sounds. These changes
consist of a larger heart rate deceleration in response to unpleasant stimuli and
higher electrodermal reactions in response to emotionally valenced stimuli (pleasant
or unpleasant) compared to neutral stimuli [92]. However, a characterization of
the affective state in terms of arousal and valence using EDA signal had not been
performed yet.
Also in this computational study, the proposal is to automatically recognize
arousal and valence levels of the standardized affective acoustic stimuli, gathered
from the IADS dataset, using only the EDA analysis. IADS sounds have been
already used in the literature (often jointly with IAPS picture) showing changes
in ANS dynamic [282, 283] and a relationship with EDA variations [284, 285].
In order to perform this research we designed a new experiment where IADS
sounds were administered to a group of young participants. During the experimental
sessions, EDA signal was continuously monitored. In this application, we demon-
strated that the convex optimization-based EDA model is suitable for affective
computing on IADS elicitation. After a monovariate and multivariate analysis,
CvxEDA, in fact, allows a better discrimination than CDA.

and Acquisition Set-Up
Twenty-five healthy subjects, aged from 25 to 35, participated as volunteers in

the experiment. According to the self-report questionnaires, none of them had a
history of injury of the auditory canal or partial or full incapability of hearing.
Moreover, none of them suffered from any mental or chronic disease. Participants
were informed about the protocol and about the purpose of the study, but they were
not informed about the arousal and valence levels they would have been listened
to. During the experiment, participants were seated in a comfortable chair in a
controlled environment while listening to the IADS sounds. Each subject was left
alone in the room where the experiment took place for the whole duration (about
29 min). The acoustic stimulation was performed by using headphones while the
subject’s eyes were closed, to avoid any kind of visual interference.
The affective elicitation was comprised of 8 sessions: after an initial resting
session of 5 min, three arousal sessions alternated with neutral sessions (see
Fig. 5.3). The three arousal levels had different increasing scores (labeled as L (low),
M (medium) and H (high)). Within each arousing session, the acoustic stimuli were
selected to have both negative and positive valence. Such levels were set according
to the IADS valence and arousal scores reported in Table 5.5. The neutral session
had a duration of 1 min and 28 s, while the three arousal sessions had a duration
of 3 min and 40 s, 4 min, and 5 min and 20 s, respectively. The different duration of
each arousal session is due to the different length of acoustic stimuli having the same
range of positive and negative valence. This experimental protocol was approved by
the local ethical committee.
During the elicitation, the EDA was continuously acquired, by means of a
dedicate hardware module of BIOPAC MP150 acquisition system.
5.4.2 Feature Extraction and Statistical Analysis
Several features were computed from the tonic and phasic components as outputs
of both the cvxEDA and CDA models. The mean value of the tonic component
was calculated in order to estimate the general psychophysiological status of the
AROUSAL VALENCE
10 2
(Pos)
0
5
(Neg)
-2
REST (N) (N) (N)
0
L M H
(AR1) (AR2) (AR3)
Fig. 5.3 Timeline of the experimental protocol in terms of arousal and valence levels. The vertical
axis relates to the IADS score, whereas the horizontal axis relates to the time. The neutral sessions,
which are marked with blue lines, alternate with the arousal ones, which are marked with red
staircases. Along the time, the red line follows the four arousal sessions having increasing intensity
of activation. The dotted green line indicates the valence levels within an arousing session. The
neutral sessions are characterized by lowest arousal and medium valence scores. Yellow line relates
to the resting state
Table 5.5 Rating of IADS sounds used in this work

Valence Arousal
Session N. of sounds rating Valence range rating Arousal range
Neutral 8 5.915˙0.68 4.346.44 3.47˙0.175 2.883.93
Arousal 1 19 / 2.467.78 5.42˙0.22 5.005.89
Arousal 2 26 / 2.047.90 6.48˙0.25 6.006.99
Arousal 3 20 / 1.577.67 7.32˙0.22 7.03 8.16
Ratings are expressed as median and its absolute deviation
Table 5.6 List of features extracted from EDA phasic and tonic compo-
nents
Feature Description
Npeak Number of significant SMNA SCR wrw
AUC Area under curve of SMNA signal wrw (S s)
Peak Maximum amplitude of significant peaks of
SMNA signal wrw3 (S)
Stdphasic Standard deviation of SMNA signal wrw (S)
MeanTonic Mean value of the tonic component
within session windows (S)
wrw within response window (i.e., 5 s after stimulus)
subjects. The characterization of the stimulus response was achieved from the sparse
phasic signal. Specifically, we calculated the number of the peaks, their maximum
amplitude and the area under curve within a time response window of 5 s after each
stimulus onset (see Table 5.6).
Table 5.7 Confusion matrix AR1 AR2 AR3

CDA features set AR1 52 28 28
AR2 16 40 28
AR3 32 32 44
Results are presented as a per-
centage value
The bold values represent
the percentage of number
of predictions/classifications
that were correct (true posi-
tives and true negatives)
Table 5.8 Confusion matrix AR1 AR2 AR3

cvxEDA features set AR1 76 16 8
AR2 12 76 12
AR3 12 8 80
centage value
the percentage of number
of predictions/classifications
that were correct (true posi-

CDA features set Positive 83 74
Negative 17 26
age value
For each feature, we statistically compared the two levels of valence, i.e.,
positive and negative, and the three arousal levels. Each statistical comparison was
performed using the nonparametric Wilcoxon tests for paired samples, given the
non-gaussianity of samples (p<0.05 from Shapiro-Wilk test with null hypothesis
of having a Gaussian sample). The multiple pairwise comparisons among the
four arousal levels were corrected with the Bonferroni’s method, multiplying the
p-values by the number of comparisons, i.e., 3. Furthermore, the multivariate pattern
analysis was performed in order to classify different levels of arousal and two levels
of valence (Tables 5.7, 5.8, 5.9, and 5.10).
Table 5.10 Confusion Positive Negative

matrix of valence levels using
the cvxEDA features set Positive 84 16
Negative 16 84
age value
1.65
2.3
1.6
2.2
mean ± SE rank
mean ± SE rank
2.1 1.55
2 1.5
1.9 1.45
1.8 1.4
1.7 1.35
A1 A2 A3 Pos Neg
Fig. 5.4 Within-subject ranks of the area under curve of the phasic component obtained by
cvxEDA for the three arousal levels (top) and the two valence levels (bottom). The dots mark
the across-subject average rank for each level while the wiskers indicate the standard error. The
three arousal levels (top) and the two valence levels (bottom) were statistical compared by the
Bonferroni-corrected Wilcoxon test that did not show significant differences
5.4.3 Experimental Results
All the features extracted from the phasic component with both methods, did not
show any significant differences among the three levels of arousal and the two
levels of valence (as example, Figs. 5.4 and 5.5 plotted the statistical description of
the AUC feature comparisons). Instead, the cvxEDA mean tonic value statistically
discriminates the three arousal and the two valence levels, showing a monotonic
trend, although inverse (Figs. 5.6 and 5.7).
5.4.3.1 Classification of Auditory Arousal and Valence Levels
Concerning the arousal levels, the multivariate pattern recognition analysis showed a
strong difference between the accuracy of the cvxEDA feature (77:33 %) (Table 5.8)
set and the CDA feature set (45:33 %) (Table 5.7). In the same way, concerning the
valence levels the convex-optimization approach showed a good performance, dis-
cerning the pleasant and unpleasant sounds with an accuracy of 84 % (Table 5.10),
whereas the CDA achieved only 54:5 % of mean successful recognition (Table 5.9).
2.15
1.6
2.1
mean ± SE rank
mean ± SE rank
2.05 1.55
2 1.5
1.95
1.45
1.9
1.4
1.85
A1 A2 A3 Pos Neg
Fig. 5.5 Within-subject ranks of the area under curve of the phasic component obtained by CDA
model for the three arousal levels (top) and the two valence levels (bottom). The dots mark
the across-subject average rank for each level while the wiskers indicate the standard error. The
three arousal levels (top) and the two valence levels (bottom) were statistical compared by the
Bonferroni-corrected Wilcoxon test that did not show significant differences
2.8
1.8
2.6
1.7
2.4
2.2 1.6
mean ± SE rank
mean ± SE rank
2 1.5
1.8
1.4
1.6
1.4 1.3
1.2 1.2
1
1.1
0.8
1
A1 A2 A3 Pos Neg
Fig. 5.6 Within-subject ranks of the mean tonic value of the cvxEDA model for the three arousal
levels (top) and the two valence levels (bottom). The dots mark the across-subject average rank for
each level while the wiskers indicate the standard error. Post-hoc Bonferroni-corrected pair-wise
comparisons of the peak amplitudes found significant differences in the cases indicated by asterisks
( W p < 0:05I W p < 0:01I W p < 0:001)
5.5 Affective Touch Elicitation
Affective touch plays an important role in social behavior and interactions [207].
Physiologically, affective haptic perception is mediated by the unmyelinated C
tactile (CT) fibers [208, 209], whose activity is linked to the contralateral primary
and bilateral secondary somatosensory area, as well as contralateral middle and
posterior insula cortex, contralateral posterior parietal cortex [210]. Since such
cortices are known to be responsible for crucial homeostatic functions involving
Autonomic Nervous System (ANS) signaling from the whole body, several previous
studies have demonstrated the correlation between ANS dynamics and affective
elicitation [156].
5.5 Affective Touch Elicitation 69
1.7
2.25
2.2 1.65
2.15 1.6
mean ± SE rank
mean ± SE rank
2.1
1.55
2.05
2 1.5
1.95
1.45
1.9
1.85 1.4
1.8 1.35
1.75
1.3
A1 A2 A3 Pos Neg
Fig. 5.7 Within-subject ranks of the mean tonic value of the CDA model for the three arousal
levels (top) and the two valence levels (bottom). The dots mark the across-subject average rank for
each level while the wiskers indicate the standard error. The three arousal levels (top) and the two
valence levels (bottom) were statistical compared by the Bonferroni-corrected Wilcoxon test that
did not show significant differences
Tactile stimulations were used in several studies where EDA was monitored
and analyzed, specifically in relation with the study of nociceptive pain. EDA
resulted useful in a fast and continuously detection of nociceptive pain with higher
sensitivity and specificity than other available objective methods [286]. Moreover,
tactile stimulation of the preterm newborn infants was found to produce significantly
higher increases in EDA than nociceptive stimulation [287].
Few studies report on the link between EDA and affective haptic stimulation.
A work by Olausson et al. [288] examined the effects on EDA dynamics of the
affective touch applying a tactile stimulation by means of a soft brush to the radial
side of the forearm. The authors performed a comparison between the effects of
pleasant touch on 4 healthy subjects and 2 patients with sensory neuronopathy
syndrome. In both patients with neuropathy, a strong sympathetic response was
reported with a latency significantly longer than in healthy subjects, suggesting a
different response between the two groups. However the small samples did not allow
a robust statistics.
In this research, we reports a case study, which is part of the European
project “WEARHAP” (WEARable HAPtics for humans and robots). During the
experimental sessions, EDA signal was continuously monitored during affective
haptic stimulation. Specifically, we had two main goals: first, to assess the physical
characteristics of the caresses (i.e. force and velocity) in terms of the arousal and
valence scores; second, we aimed to process the EDA in order to extract features
correlated to force and velocity levels of affective touch stimulation and perform an
automatic affective recognition.
In this view, we used a haptic device able to mimic the human caress [247]
(Fig. 5.8). Specifically, a layer of fabric is stretched by two motors, whose position
and velocity determine the force and velocity of the simulated caress. In this way, the
artificial caresses can be controlled and standardized in terms of force and velocity
[247].
Fig. 5.8 An overview of the

haptic system worn by a
subject
Previous works have already investigated about the perception of affective touch
as a function of the velocity and force of the stimulus. As an example, a peak
in pleasantness perception was found for stimuli having velocity of 3 cm/s [211].
Of note, this velocity was found to be optimal for the activation of CT fibers
[289]. Moreover, low force of caressing was considered more pleasant [212]. Other
physical parameters, e.g., roughness, were found to affect the affective perception
of the tactile stimulus [290].
5.5.1 A Device for Caress-Like Haptic Stimuli
In this research the haptic stimuli were reproduced by means of an affective touch
display [247] that simulates the human caress.
The device uses a layer of elastic fabric. After having tested several materials
(including commercial lycra, latex layer, and silicon rubber), we selected the elastic
Superbiflex HN by Mectex S.P.A for its high resistance to traction and the good
elastic behavior exhibited in a large range of elasticity.
More specifically, the extremities of a rectangular-shaped fabric (60 mm
160 mm) are connected by means of screws to two rolls, each of them can be
independently moved by one motor (HITEC digital DC servomotor HS-7954 H
with an input voltage of 7.4 V). The choice of this kind of motors is motivated by the
fact that they provide a good trade-off between velocity and torque, thus enabling
fast changes in the stretching state of the fabric. The motors are also suitably
modified to allow a continuous rotation of the motor shafts, which are connected
directly to the rolls (see Fig. 5.9 for further details). The motor positions and rotation
velocity are controlled by a customized electronic board (PSoC-based electronic
board with RS–485 communication protocol), which reads motor positions by using
two absolute magnetic encoders (12 bit magnetic encoder by Austrian Microsystems
AS5045 with a resolution of 0:0875ı ).
Finally, the system is endowed with a load cell (Micro Load Cell [0–5 kg] with
a resolution of 5 g by Phidgets) placed at the basis of the forearm support to record
Fig. 5.9 The system: exploded draw without the cover. Each component of the system is reported.
The total dimensions of the system can be inscribed in a parallelepiped of dimensions 150 150
80 mm
the normal force exerted by the fabric on the forearm. The device was realized in
ABS plastic material and it is encapsulated in a plastic cover to protect motors.
The system exploded draw is reported in Fig. 5.9.
The system is designed to be wearable, as it can be applied to the forearm of a
subject without discomfort (see also Fig. 5.8). The subject places her/his forearm
on the forearm support (half hollow cylinder, radius 45 mm and length 60 mm, see
Fig. 5.9), under the fabric. When the device is operating, two different phases can
be individuated (see Fig. 5.10):
Calibration Phase: in this phase it is possible to calibrate the force exerted
by the fabric on the subject forearm. There are two modalities of control: when
motor 1 rotates in a counter-clockwise direction and motor 2 rotates in a clockwise
direction, the fabric is increasingly wrapped around the forearm thus pressing it with
an increasing force (as it is reported in Fig. 5.10), that can be recorded by the load
cell. Conversely, when motor 1 starts to rotate in a clockwise direction and motor
2 in a counter-clockwise direction the force exerted by the fabric on the forearm
decreases. In this manner it is possible to modulate the “strength of the caress” by
suitably controlling the positions of the motors. The maximum level of force that
can be applied is 20 N.
Movement Phase: when the desired level of force exerted by the fabric is
achieved and both motors are in the reference positions, they start to coherently
Fig. 5.10 The scheme of the working principle of the system. The Movement Phase is displayed
in green while the Calibration Phase is displayed in blue: in the example of the figure, the level
of force exerted by the fabric must be increased to reach the desired one. The radius of each roll
(15 mm) is reported
rotate and the fabric moves forward and backward over the subject forearm,
simulating a caress. The velocity of the caress can be modulated by regulating the
velocity of the motors. More specifically, since the electronic board is endowed with
a built-in motor position controller, we can feed the motors with a sinusoidal input
reference trajectory, whose frequency and amplitude can be set, thus controlling the
velocity and the amplitude of motor rotation, respectively. The maximum angular
displacement of the motors from the reference positions is set to ˙90ı . An entire
control cycle lasts 1 ms and the fastest sinusoidal input (frequency of 5 Hz) can be
followed with a phase delay less than 2 ms.

and Acquisition Set-Up
Thirty-two healthy subjects aged 27 ˙ 2 (16 males, age 26 ˙ 2, and 16 females,

age 27 ˙ 2) gave their informed consent to take part in the study. This study
was approved by the Ethical Committee. Participants were comfortably seated with
N
mm
s
0.7 6
0.4 Velocity
Force
2
0.1
Initial Rest Final Rest

3 min Touch ~12 min
Stimuli Time
Fig. 5.11 Example of protocol scheme of one subject. Velocity and force combination is
randomized
the right forearm horizontally placed on the support of the haptic device. For all
trials, participants wore earplugs in order to prevent any auditory cues. Six different
combinations of stimuli among 2 levels of force (2 N, 6 N) and 3 levels of velocity
(9.4, 37, and 65 mm/s) were used. During the interval between two stimuli, the
motors were stopped and the force was set to 0 N, in this case the fabric was only
lightly in contact with the forearm.
The length of the experiment was divided into 3 consecutive sessions (see
Fig. 5.11):
• resting state session of 3 min;
• automatic caress session of 12 min: the participants are subject to haptic
stimulations by the affective touch display on the right forearm. All the 6
combinations of stimuli (two force and three velocity) are administered;
• final resting state session of 2 min.
Of note, all the six combinations of velocities and forces were suitably random-
ized among subjects, with a pre-stimulus and a post-stimulus interval of 35 s each.
After each post-stimulus, the subjects were asked to assess the stimulus in terms of
arousal and valence scores [291], within a time window of 20 s. The evaluation
of the emotions elicited is performed adopting the simplified version of CMA
[247, 248, 292] that allows taking into account the two main dimensions: valence
and arousal.
Table 5.11 List of features extracted from EDA phasic and tonic
components
Phasic feature Description
Npeak Number of significant SCR wrw
AUC Area under curve of SMNA signal wrw (S s)
Stdphasic Standard deviation of SMNA signal wrw (S
Tonic feature Description
MeanTonic Mean value of the tonic component wrw (S)
MeanTonic difference between
diffTonic post/pre rest sessions (S)
Difference of AUC of spontaneous SCRs
NsAUC between post/pre rest sessions ( S s)
wrw within response window (i.e., 5 s after stimulus)
5.5.3 Feature Extraction, Performance Metrics,

and Statistical Analysis
Once the tonic and phasic components were estimated from both cvxEDA and
CDA approaches, several features could be extracted to investigate the sensitivity
to changes in caressing, along the force and velocity dimensions. Proposed features
in this study are summarized in Table 5.11.
Through these features, the following analyses were performed:
• Event-related phasic analysis, through which EDA was studied within a time
response window of 5 s after the affective stimulus;
• Non specific fluctuation and tonic analysis, comprised of (i) the tonic value,
averaged within the time-response window, and the nonspecific electrodermal
fluctuations, and (ii) differential tonic value between the post- and pre-stimulus
session.
The differences between the two levels of force (i.e., F1 D 2 N and F2 D 6 N) and
among the three velocities (i.e., V1D9.4 mm/s, V2D37 mm/s and V3D65 mm/s)
were studied using Wilcoxon signed-rank tests and Friedman tests, respectively,
due to the non-Gaussianity of samples (as confirmed by a preliminary analysis
performed through Shapiro-Wilk tests). In case of rejection of the Friedman test
null-hypothesis, a post-hoc analysis, using Wilcoxon signed-rank with Bonferroni
correction, was also carried out.
Finally, to test the suitability of the proposed algorithm in embedded computing
systems, computational performance analysis was also performed. This is related to
the execution time of the algorithm (processing time), and the memory usage of the
CDA and cvxEDA models output vectors. Estimates of these metrics were obtained
on the same personal computer, with processor 1.7 GHz Intel Core i7, RAM memory
8 GB 1600 MHz DDR3.
Arousal−Velocity Arousal−Force
p < 0.02 p < 0.001
2.5 1.9
1.8
1.7
Ranks
Ranks
1.6
2
1.5
1.4
1.3
1.5
9.4 37 65 2 6
Velocity (mm/s) Force (N)
Valence−Velocity Valence−Force
p < 1e−06 p < 0.005
1.9
2.6 1.8
2.4 1.7
Ranks
1.6
Ranks
2.2
2 1.5
1.8 1.4
1.6 1.3
1.4
9.4 37 65 2 6
Velocity (mm/s) Force (N)
Fig. 5.12 Multiple comparisons of Arousal (top figures) and Valence (bottom figures) values
correspondent to three levels of velocities (9.4, 37, and 65 mm/s) and two levels of forces (F D 2 N,
F D 4 N)
5.5.4 Statistical Results of the Self Assessment Questionnaire
Results from the self-assessment-questionnaire [291] are summarized in Fig. 5.12 as

boxplots. Specifically, non-parametric statistics revealed that the lower force level
(2 N) was associated to a pleasant and low arousing elicitation (p<0.01).
Concerning the velocity of the caress, it was possible to associate the stimulus
with higher velocity (V3) to the one with higher arousal and lower valence
(unpleasant). In particular, the post-hoc multiple pairwise comparisons showed
that only the arousal scores associated to V1 and V3 were statistically different
(p < 0:01). Instead, concerning the valence, V1 scores were significant higher than
V2 and V3 (p < 105 ), which, however, were statistically indistinguishable from
each other.
5.5.5 Experimental Results of Tactile Stimulation
In this section, results from the statistical analysis performed on features from CDA
and cvxEDA algorithms are reported. Such results are shown considering EDA non-
specific fluctuations and tonic components analysis, as well as event-related phasic
components analysis (Fig. 5.13).
5.5.5.1 EDA Non-specific Fluctuations and Tonic Components Analysis
Only one feature, diffTonic, extracted from both the CDA and cvxEDA models
showed a significant difference (p < 0:005, Fig. 5.14) between the two levels
of force while caressing. In particular, the higher diffTonic values, the higher
the caressing force. Note that diffTonic calculated through cvxEDA showed a
higher discerning power, i.e., lower p-value, as compared with CDA modeling (see
Tables 5.12 and 5.13).
Concerning the three levels of caress velocity, among the features extracted using
the CDA model, significant differences (p < 0:05) were found in diffTonic while
discerning V1 vs. V2, and V1 vs. V3, and in diffNSAUC while discerning V1 vs.
V3 and V2 vs. V3 (see Table 5.14 and Figs. 5.16 and 5.17). On the other hand,
using the cvxEDA model, a significant difference in discerning V1 vs. V2 was found
through meanTonic (see Table 5.15 and Fig. 5.16). CDA showed a higher number of
statistical differences among the velocities through the tonic features than cvxEDA,
though a complete discrimination of all of the three levels is not achieved by both.
5.5.5.2 EDA Event-Related Phasic Components Analysis
This analysis revealed no significant differences between the two caressing forces,
being consistent between the CDA and cvxEDA approaches (see Tables 5.12
and 5.13, and Fig. 5.15).
Concerning differences between the three caressing velocities, CDA modeling
showed significant differences (p < 0:002) on the number of significant peaks
(Npeak), and the area under the phasic signal curve (AUC) while discerning V1
vs. V2, and V1 vs. V3. No significant differences were found otherwise, including
V2 vs. V3 (see Table 5.14 and Fig. 5.17). Using the cvxEDA approach, instead,
we found significant differences in all of the pairwise comparisons, with p < 0:005
(see Table 5.15 and Fig. 5.17). Importantly, a monotonically increasing trend among
the three velocities was found for each of the considered features (see Fig. 5.17).
Figure 5.13 shows exemplary EDA phasic responses from CDA and cvxEDA
models, for each combination of caressing force and velocity. Note that, considering
sparse outputs from cvxEDA, amplitude values are consistent with the caressing
velocity level.
Stimulus CDA model cvxEDA model

−3
x 10
14 10
12 8
10
6
[a.u.]
[a.u.]
8
4
6
4 2
F1, V1 2
0 10 20
t [s]
30 40
0
0 10 20
t [s]
30 40
−3
x 10
12 12
10 10
8
8
[a.u.]
[a.u.]
6
6
4
4 2
F1, V2 2
0 2 4 6 8
0
0 2 4 6 8
t [s] t [s]
0.035 15
0.03
0.025 10
[a.u.]
[a.u.]
0.02
0.015 5
0.01
F1, V3 0.005
0 1 2
t [s]
3 4 5
0
0 1 2
t [s]
3 4 5
0.02 6
5
0.015
4
[a.u.]
[a.u.]
0.01 3
2
0.005
1
F2, V1 0
0 10 20
t [s]
30 40
0
0 10 20
t [s]
30 40
0.025 15
0.02
10
0.015
[a.u.]
[a.u.]
0.01
5
0.005
F2, V2 0
0 2 4
t [s]
6 8
0
0 2 4
t [s]
6 8
0.035 15
0.03
10
0.025
[a.u.]
[a.u.]
0.02
5
0.015
F2, V3 0.01
0 1 2
t [s]
3 4 5
0
0 1 2
t [s]
3 4 5
Fig. 5.13 Example of CDA (left) and cvxEDA (right) phasic responses to each combination
of caressing force (F D 2 N, F D 6 N) and velocity (V1D 9.4 mm/s, V2 D 37 mm/s and
V3D 65 mm/s) level, within the stimulus time window a.u. stands for arbitrary unit.
CDA model cvxEDA model
meanTonic meanTonic
1.7 1.7
mean ± SE rank
mean ± SE rank
1.6 1.6
1.5 1.5
1.4 1.4
1.3 1.3
F1 F2 F1 F2
diffTonic diffTonic
1.8
mean ± SE rank
mean ± SE rank
1.6 1.6
1.4 1.4
1.2 1.2
F1 F2 F1 F2
diffNSAUC diffNSAUC
mean ± SE rank
mean ± SE rank
1.7
1.6 1.6
1.5 1.5
1.4 1.4
1.3
F1 F2 F1 F2
Fig. 5.14 Within-subject ranks of the tonic feature set obtained from CDA (left) and cvxEDA
(right) models between the two force levels (F1D 2 N, F2D 6 N). Values represent average rank
˙ standard error (SE) across subjects. Asterisks indicate significant differences between velocities:
./p < 0:05; ./p < 0:01; . /p < 0:001
Table 5.12 Median ˙ MAD (median absolute deviation) interval for

CDA features
Feature F1 F2 p-value
Npeak 0˙0 0˙0 0:690
Peak 0.0419 ˙ 0.0255 0.0387 ˙ 0.0256 0:653
AUC 0.236 ˙ 0.176 0.263 ˙ 0.202 0:256
Stdphasic 0.0108 ˙ 0.00891 0.00996 ˙ 0.00834 0:855
MeanTonic 1.67 ˙ 0.452 1.55 ˙ 0.413 0:262
diffTonic 0.0457 ˙ 0.067 0.0113 ˙ 0.0624 0:00354
diffNSAUC 0.112 ˙ 0.122 0.151 ˙ 0.156 0:186
Values were averaged among the subjects. Last column shows p-values
from Wilcoxon non-parametric tests, with null hypothesis of equal
median values between two force levels
Table 5.13 Median ˙ MAD Feature F1 F2 p-value

intervals for cvxEDA features
Npeak 3.0 ˙ 1.0 3.0 ˙ 1.0 0:343
Peak 3.99 ˙ 3.94 3.78 ˙ 3.58 0:919
AUC 24.6 ˙ 21.6 32.0 ˙ 27.5 0:338
Stdphasic 0.414 ˙ 0.41 0.412 ˙ 0.378 0:913
MeanTonic 0.289 ˙ 0.476 0.406 ˙ 0.445 0:225
diffTonic 0.109 ˙ 0.231 0.0453 ˙ 0.382 1.26e-4
diffNSAUC 2.11 ˙ 7.01 4.9 ˙ 10.1 0:286
Values were averaged among the subjects. Last column shows p-
values from Wilcoxon non-parametric tests, with null hypothesis
of equal median values between two force levels
5.5.5.3 Classification of Haptic Force and Velocity Levels
Tables 5.16 and 5.17 show the confusion matrix relative to the classification process
of the three velocity levels. In this case, even if the statistical results showed a
better discrimination for the cvxEDA features set, the CDA features have a better
accuracy (67:7 %). Otherwise, concerning the force level classification, the accuracy
of the cvxEDA (68:75 %) model was more than 10 % compared to CDA (58:33 %)
(Tables 5.18 and 5.19).
Independently form the model used in the analysis, these no satisfactory results
of the classification procedure suggested the idea to perform a classification in
the arousal and valence dimensions without taking into account the physical
characteristics of the stimulus.
5.5.5.4 Classification of Arousal and Valence Levels

After Haptic Elicitation
The arousal and valence scores assigned by the recruited subjects were grouped
into two main classes (Tables 5.20, 5.21, 5.22 and 5.23). Concerning the arousal
levels, we identified a neutral class (arousal < 2), and an aroused class (arousal > 2).
Concerning the valence levels we divided the stimuli in a pleasant class (valence >
0), and in an unpleasant class (valence < 0), In this case due to the unpaired nature of
the samples, a SVM classifier was applied to the dataset. The confusion matrixes for
both the arousal and valence problems showed a better recognition for the cvxEDA
method. The accuracy of the arousal classification was 75:14 % (Table 5.21) against
63:35 % of the CDA (Table 5.20). Regarding the valence classification, the cvxEDA
reached an accuracy of 77:96 % (Table 5.23) whereas the CDA model showed only
66:11 % (Table 5.22).
80
Table 5.14 Median ˙ MAD intervals for CDA features

p-value
Feature V1 V2 V3 V1-V2 V1-V3 V2-V3
Npeak 0˙0 1.0 ˙ 1.0 1.0 ˙ 1.0 0:00144 6.33e-6 0:296
peak 0.0519 ˙ 0.0308 0.0344 ˙ 0.0216 0.0397 ˙ 0.0255 1 0:132 0:226
AUC 0.0964 ˙ 0.0634 0.331 ˙ 0.241 0.343 ˙ 0.254 8.12e-7 1.852e-9 0:164
Stdphasic 0.0134 ˙ 0.0108 0.0103 ˙ 0.0076 0.00924 ˙ 0.00773 0:678 0:062 0:75
MeanTonic 1.53 ˙ 0.369 1.69 ˙ 0.467 1.64 ˙ 0.463 0:0599 0:505 1
diffTonic 0.0663 ˙ 0.0815 0.00352 ˙ 0.0773 0.0175 ˙ 0.0599 0:00172 0:0393 1
diffNSAUC 0.0633 ˙ 0.0697 0.127 ˙ 0.148 0.259 ˙ 0.219 0:511 0:00286 7.05e-4
Values were averaged among the subjects. Last three columns shows p-values from Wilcoxon non-parametric tests,
with null hypothesis of equal median values between three velocity levels
5 Experimental Applications on Multi-Sensory Affective Stimulation
Table 5.15 Median ˙ MAD intervals for cvxEDA features
5.5 Affective Touch Elicitation
p-value
Feature V1 V2 V3 V1-V2 V1-V3 V2-V3
Npeak 2.0 ˙ 1.0 3.0 ˙ 1.5 4.0 ˙ 2.0 0:00409 5.36e-07 2.82e-04
Peak 0.256 ˙ 0.256 4.36 ˙ 4.03 7.61 ˙ 6.16 8.55e-06 5.68e-07 0:261
AUC 6.89 ˙ 6.59 32.4 ˙ 24.8 72.6 ˙ 42.0 7.61e-08 1.11e-09 1.34e-05
Stdphasic 0.0356 ˙ 0.0355 0.469 ˙ 0.406 0.752 ˙ 0.577 4.25e-07 2.01e-08 0:0215
MeanTonic 0.527 ˙ 0.471 0.261 ˙ 0.429 0.292 ˙ 0.568 0:0364 0:111 1
diffTonic 0.122 ˙ 0.335 0.0251 ˙ 0.303 0.0417 ˙ 0.221 0:0689 0:0666 1
diffNSAUC 2.1 ˙ 6.61 2.13 ˙ 9.22 5.43 ˙ 9.37 1 1 0:344
Values were averaged among the subjects. Last three columns shows p-values from Wilcoxon non-parametric
tests, with null hypothesis of equal median values between three velocity levels
81
Npeak Npeak
1.7
mean ± SE rank
mean ± SE rank
1.6
1.6
1.5 1.5
1.4
1.4
1.3
F1 F2 F1 F2
peak peak
1.7
mean ± SE rank
mean ± SE rank
1.6 1.6
1.5 1.5
1.4 1.4
1.3
F1 F2 F1 F2
AUC AUC
mean ± SE rank
mean ± SE rank
1.7 1.7
1.6 1.6
1.5 1.5
1.4 1.4
1.3 1.3
F1 F2 F1 F2
stdphasic stdphasic
mean ± SE rank
mean ± SE rank
1.6 1.6
1.5 1.5
1.4 1.4
F1 F2 F1 F2
Fig. 5.15 Within-subject ranks of the phasic feature set obtained from CDA (left) and cvxEDA
(right) models between the two force levels (F1=2 N, F2=6 N). Values represent average rank ˙
standard error (SE) across subjects. Asterisks indicate significant differences between velocities:
./p < 0:05; ./p < 0:01; . /p < 0:001
5.5.5.5 Analysis of Computational Performance
Results from the computational performance analysis, showing group-wise statis-

tics, are reported in Table 5.24.
Concerning the processing time, no significant difference was found between
the two models. However, as we show in Fig. 5.18, CDA execution time is not
linearly related to the length of the input signal. In particular, for input signals longer
5.6 Affective Olfactory Elicitation 83
meanTonic meanTonic
mean ± SE rank
mean ± SE rank
2.2 2.2
2 2
1.8
1.8
1.6
V1 V2 V3 V1 V2 V3
diffTonic diffTonic
2.4
mean ± SE rank
mean ± SE rank
2.2
2.2
2
1.8 2
1.6 1.8
1.4
1.6
V1 V2 V3 V1 V2 V3
diffNSAUC diffNSAUC
mean ± SE rank
mean ± SE rank
2.5 2.2
2 2
1.5 1.8
V1 V2 V3 V1 V2 V3
Fig. 5.16 Within-subject ranks of the tonic feature set obtained from CDA (left) and cvxEDA
(right) models between the three velocity levels (V1D9.4 mm/s, V2D37 mm/s and V3D65 mm/s).
Values represent average rank ˙ standard error (SE) across subjects. Asterisks indicate significant
differences between velocities: ./p < 0:05; ./p < 0:01; . /p < 0:001
than 1300 s, CDA processing time tends to grow superlinearly. Concerning memory
usage, a significant difference was found between the two models. As expected,
given to the intrinsic sparse nature of cvxEDA phasic components, lower storage
values were associated with the cvxEDA model.
5.6 Affective Olfactory Elicitation
Emotional experience and hedonic judgment are principal aspects of the olfactory
sense [293]. The popular belief reports the ability of fragrances to affect emotional
states. In fact, odor perception can influence our daily life in many ways such as
by modulating our behavior, our autonomic nervous system parameters, and our
cerebral activity [294–296].
Npeak Npeak
2.4
mean ± SE rank
mean ± SE rank
2.5
2.2
2 2
1.8
1.6 1.5
1.4
1.2 1
V1 V2 V3 V1 V2 V3
peak peak
2.4
mean ± SE rank
mean ± SE rank
2.5
2.2
2
2
1.8 1.5
1.6 1
V1 V2 V3 V1 V2 V3
AUC AUC
mean ± SE rank
mean ± SE rank
2.5 2.5
2 2
1.5 1.5
1 1
V1 V2 V3 V1 V2 V3
stdphasic stdphasic
2.4
mean ± SE rank
mean ± SE rank
2.5
2.2
2
2
1.5
1.8 1
V1 V2 V3 V1 V2 V3
Fig. 5.17 Within-subject ranks of the phasic feature set obtained from CDA (left) and cvxEDA
(right) models between the three velocity levels (V1D9.4 mm/s, V2D37 mm/s and V3D65 mm/s).
Values represent average rank ˙ standard error (SE) across subjects. Asterisks indicate significant
differences between velocities: ./p < 0:05; ./p < 0:01; . /p < 0:001
The close relationship between olfaction and emotion is a logical consequence

of the sharing of several limbic regions by these processes [294].
The olfactory system links structures of the limbic system belonging to the
central nervous system, specifically the amygdala and hippocampus, which are
directly involved in the modulation of emotions [294]. These anatomical connec-
tions demonstrate how the olfactory experience is inextricably linked with affect
experience [297, 298].
Table 5.16 Confusion V1 V2 V3

matrix of velocity levels
using the CDA features set V1 79.69 15.63 4.69
V2 15.63 56.25 28.13
V3 4.69 28.13 67.19
Results are presented as a
percentage value
the percentage of number of
predictions/classifications that
were correct (true positives
and true negatives)
Table 5.17 Confusion V1 V2 V3

matrix of velocity levels using
the cvxEDA features set V1 54.69 17.19 28.13
V2 25.56 46.88 18.753
V3 18.75 35.94 53.139
centage value
The bold values represent the
percentage of number of predic-
tions/classifications that were
correct (true positives and true
negatives)
Table 5.18 Confusion F1 F2

matrix of force levels using
the CDA features set F1 58.33 41.67
F2 41.67 58.33
Results are presented
as a percentage value
The bold values re-
present the percent-
age of number of
predictions/classifica-
tions that were correct
(true positives and
true negatives)
The brain limbic areas activated by a hedonic olfactory stimulus are also known
to be responsible for crucial homeostatic functions of the whole body involving
the autonomic nervous system (ANS) activity. More specifically, prior state of
the art reports a strong correlation between ANS dynamics, as estimated through
electrodermal activity (EDA) processing, and affective elicitations [57, 156].
The ANS controls heart rate variability (HRV) and EDA [135, 160, 299], which
are modulated by the perception of an odorant [250, 300–302].
In particular, EDA was found to be associated with the odorant concentration:
weak concentrations of odorants evoked lower EDA response than higher con-
centrations [303]. The correlation between odor intensity, arousal, hedonic tone
Table 5.19 Confusion F1 F2

matrix of force levels using
the cvxEDA features set F1 68.75 31.25
F2 31.25 68.75
Results are presented
as a percentage value
The bold values re-
present the percent-
age of number of
predictions/classifica-
tions that were correct
(true positives and
true negatives)
Table 5.20 Confusion Neutral Aroused

matrix of arousal levels using
the CDA features set Neutral 68.06 35.83
Aroused 31.94 64.17
centage value
tions/ classifications that were
negatives)
Table 5.21 Confusion Neutral Aroused

matrix of arousal levels using
the cvxEDA features set Neutral 77.78 27.50
Aroused 22.22 72.50
centage value
tions/ classifications that were
negatives)
Table 5.22 Confusion Pleasant Unpleasant

the CDA features set Pleasant 71.23 44.54
Unpleasant 28.77 55.46
value
The bold values represent the percent-
age of number of predictions/ classifi-
cations that were correct (true positives
and true negatives)
Table 5.23 Confusion Pleasant Unpleasant

the cvxEDA features set Pleasant 79.45 23.53
Unpleasant 20.55 76.47
value
age of number of predictions/ classifi-
and true negatives)
Table 5.24 Median ˙ MAD intervals for CDA and cvxEDA perfor-
mances
CDA cvxEDA p-value
Processing time [s] 8.171 ˙ 1.239 7.038 ˙ 1.657 0.550
Memory [Kb] 359.004 ˙ 2.868 136.65 ˙ 95.384 2.163e-4
p-values are gathered from the Mann-Whitney non-parametric tests with
null hypothesis of equal medians between models.
Values were calculated for each recording, and averaged among the
subjects
The bold value represents the significant statistical value
and familiarity has been already addressed in the literature. Henion et al. [304]
considered the intensity and hedonic tone as a single feature, whereas other authors
did not shared this idea [305–307]. Other studies showed that the EDA signal
could be modulated by odor intensity, valence, arousal or familiarity [250, 308].
Brauchli et al. also showed that the mean tonic value varied according to the
smell pleasantness, but not to arousal [302]. Findings concerning HRV are similar:
generally, unpleasant smells evoke an increase of the mean HRV value, and
viceversa [295, 302, 308].
Several papers reported gender differences in the emotion area [309–313]. There-
fore, gender differences should be taken into account when emotional paradigms
are used [314]. However, possible gender differences in physiological responses
to odorants have been rarely studied. In childhood, gender differences were found
only in response to unpleasant odors [315]. Yousem et al. [316] examined the gender
effects on odor-stimulated fMRI and evidenced a greater fMRI activation in women
than in men.
Considering the evidences of the relationship among olfactory emotion stimuli
and ANS, we analyzed the effect of olfactory stimulation on EDA. In addition to
the automatic pattern recognition system to classify the valence level of affective
olfactory stimuli, and the gender effect was investigated.
Considering these evidences of the relationship among olfactory emotion stimuli
and ANS, we studied the characterization of the physiological response to olfactory
affective stimuli in terms of arousal and valence analyzing the EDA variations.
20
CvxEDA
18
CDA
16
14
Processing Time [s]
12
10
0
0 200 400 600 800 1000 1200 1400 1600 1800
Time Window lenght [s]
Fig. 5.18 Processing time of the CDA and cvxEDA algorithms at different length of signal input,
with sampling time of 60 s

and Acquisition Set-Up of the Olfactory Stimulation
In this experiment we used five different smells, which were synthesized in the
laboratory. These odorants were selected due to their different hedonic tone and
their safety for the panel members [317].
• O1 D Vanillin (C8 H8 O3 , 152:15 g=mol)
• O2 D Benzaldehyde (C6 H5 CHO, 106:12 g=mol)
• O3 DN-butanol (CH3 CH2 CH2 CH2 OH, 74:12 g=mol)
• O4 D Isovaleric acid ( .CH3/2 CHCH2 COOH, 102:13 g=mol)
• O5 D Butyric acid (CH3 CH2 CH2 CO2 H, 88:11 g=mol)
The first two smells were considered pleasant, the last two unpleasant, a priori.
The N-butanol was considered neutral. The solutions with the specific smell were
obtained by mixing each of the chemical compounds with 500 mL of distilled water.
The final concentration of all solutions was such as to appear isointense.
Thirty-two subjects were enrolled in the experiment. In order to have the most
homogeneous subject sample as possible, we determined the olfactory perception
threshold of each participant. We preprepared ten solutions of each smell with a

different dilution factor. For example, concerning the N-butanol:
1. dilution: 99.90234375 g of distilled water C 0.024414062 g of solution
N-butanol a maximum concentration
2. dilution: 99.8046875 g of distilled water C 0.048828125 g of solution
3. dilution: 99.619375 g of distilled water C 0.09765625 g of solution of
4. dilution: 99.21875 g of distilled water C 0.1953125 g of solution of N-butanol
at Maximum concentration!
5. dilution: 98.4375 g of distilled water C 0.380625 g of solution of N-butanol at
Maximum concentration
6. dilution: 96 875 g of distilled water C 0.78125 g of solution of N-butanol at
maximum concentration
9. dilution: 75 g of distilled water C 6:25 g of solution of N-butanol at maximum
concentration
10. dilution: 50 g of distilled water C 12.5 g of solution of N-butanol at maximum
concentration
The threshold of absolute subjective perception was found by presenting to each
subject two bottles: one containing a diluted solution of N-butanol and the other
the distilled water. The procedure started testing the lower concentration up to the
concentration that the subject was able to distinguish from the water. At each step
was asked the subject to indicate if one of the two bottles had a different smell from
the water, and possibly indicate which one was perceived different. Any lack of or
incorrect answer was followed by a next test of the solution of N-butanol at higher
concentration. On the other hand, after each correct answer the previous step at the
same dilution level was repeated again. The test was considered concluded when
the subject gave four subsequent correct answers. To avoid olfactory adaptation, we
expected a rest session of 45 s after each step. We selected only the subjects with the
olfactory perception threshold of 4.
The experimental timeline was designed as follows (see Fig. 5.19):
• 3 min of initial rest;
• 1 min of pre-stimulus rest;
• 5 s of the first olfactory stimulation;
• 1 min of post-stimulus rest;
• 20 s for the self assessment questionnaire (SAM);
• 1 min of pre-stimulus rest;
• 5 s of the second olfactory stimulation;
• 1 min of post-stimulus rest;
Fig. 5.19 Timeline of the experimental protocol: each stimulus was randomized and between two
resting sessions. After the post-stimulus rest the subject scored the stimulus in terms of arousal and
valence levels
• 20 s for the self assessment questionnaire;

• 3 min of final rest;
After each stimulus the participant was ask to score the stimulus in terms of
arousal (with a scale from 1 to 5) and valence (with a scale from 2 to 2) using the
SAM technique [291].
5.6.2 Feature Extraction and Statistical Analysis
Following the same approach used in the affective touch protocol, we extracted
the features as summarized in Table 5.35. Likewise the affective touch, a statistical
analysis among the five smells were performed on the arousal and valence scores,
and on the EDA phasic and tonic features. We distinguished the event-related phasic
analysis, i.e., EDA was studied within a time window of 5 s correspondent to
the affective stimulus session, and the non specific fluctuation and tonic analysis,
comprising the tonic level comparison and the differential value of the tonic features
between the post- and pre-stimulus resting session.
The differences among the smells were studied using the Friedman tests, and
in case of rejection of the Friedman test null-hypothesis, a post-hoc analysis was
performed by means of a Bonferroni corrected Wilcoxon signed-rank.
5.6.3 Statistical Analysis on Self-Assessment

Questionnaire Scores
Results from the self-assessment-questionnaire are shown in Fig. 5.20. No signifi-

cant differences were found among the arousal scores. Concerning the pleasantness,
it is possible to identify three groups : the N-butanol was the neutral smell, the
Vanillin and Benzaldehyde were assessed as pleasant and the Isovaleric and Butyric
AROUSAL VALENCE
3.2 1.5
3.1 1
3
0.5
mean ± SE rank
mean ± SE rank
2.9
0
2.8
−0.5
2.7
−1
2.6
2.5 −1.5
2.4 −2
2.3 −2.5
O1 O2 O3 O4 O5 O1 O2 O3 O4 O5
Fig. 5.20 Multiple comparisons of Arousal (left figure) and Valence (right figure) values corre-
spondent to the five smells. Legend: O1 D Vanillin; O2 D Benzaldehyde; O3 DN-butanol;
O4 D Isovaleric acid; O5 D Butyric acid. Asterisks indicate significant differences between
smells: ./p < 0:05; ./p < 0:01; . /p < 0:001
Table 5.25 Confusion Pleasant smell Unpleasant smell

the CDA features set Pleasant smell 68.75 31.25
Unpleasant smell 31.25 68.75
Results are presented as a percentage value
The bold values represent the percentage of number of
predictions/classifications that were correct (true posi-
acid were the unpleasant smells. In conclusion, each smell elicited the same intensity
sensation but confirmed the a priori consideration regarding the valence level.
The subsequent EDA analysis was based on the SAM results in order to find a
relationship with the response to the pleasant and the unpleasant stimuli.
5.6.4 Statistical Analysis and Classification of Olfactory

Valence Levels
The statistical analysis of all the feature extracted from both CDA and cvxEDA did
not show any significant differences among the smells. Neither features extracted
from the phasic or the tonic components were not able to find any differences in the
relationship between EDA and the five odors. However, considering the multivariate
analysis, we grouped the two positive and two negative smells, without taking
into account the arousal dimension due to the fact that no statistical differences
were found in the SAM analysis. Both method showed the same recognition
performances. CDA and cvxEDA features were able to classify the pleasant and
unpleasant stimulus with an accuracy of 68.75 % (see Tables 5.25 and 5.26).
Table 5.26 Confusion Pleasant smell Unpleasant smell

the cvxEDA features set Pleasant smell 68.75 31.25
Unpleasant smell 31.25 68.75
predictions/classifications that were correct (true posi-
5.6.5 Dataset Reduction and Gender Analysis
Considering the results obtained in the classification and in the SAM analyses, we
decided to reduce the stimulus set to two odorants. According to the SAM results
and the literature review [317], we selected the smells with the two intermediate
median levels of arousal among five smells of the previous analysis.
O1 : Benzaldehyde C6 H5 CHO, (concentration 106:12 g=mol);
O2 : Isovaleric acid .CH3/2 CHCH2 COOH, (concentration 102:13 g=mol).
The described classification procedure for paired data was applied for valence
recognition of the three following datasets (outputs of the cvxEDA model):
x: the whole subject dataset,
z: the reduced dataset of all the male subjects
w: the reduced dataset of all the female subjects
Considering the three datasets, results from the self-assessment-questionnaire are
shown in Fig. 5.21. No significant differences were found among the arousal scores
in the three datasets (x;y;z). Concerning the valence dimension, we can statistically
discern a pleasant smell and an unpleasant smell in the three groups of participants
(p < 106 ). Both smells elicited the same intensity sensation but confirmed the
a priori consideration regarding their different valence level, both in males and
females.
Results of the classification procedure on the three datasets are shown in the form
of a confusion matrix in Table 5.27. Considering data from all of the subjects (x),
i.e., men and women, cvxEDA features were able to recognize the two valence levels
with an average accuracy of 68.76 %. Given the poor classification performance, we
hypothesized that gender could have significantly affected the system accuracy, and
then split the dataset into two sub-sets according to gender. With the male dataset
(z), recognition accuracy was still poor 62.5 %, whereas an accuracy of 78.13 % was
achieved with data from females (w).
5.7 Assessment of Mood States in Bipolar Patients Using EDA 93
AROUSAL
AROUSAL AROUSAL
5
4 4
3.5 3.5
4
Mean ± STD
Mean ± STD
Mean ± STD
3 3
2.5 2.5 3
2 2
2
1.5 1.5
1 1 1
O1 O2 O1 O2 O1 O2
VALENCE VALENCE VALENCE
3 3 2
2 2 1
Mean ± STD
Mean ± STD
Mean ± STD
1
1 0
0
0 −1
−1
−2 −1 −2
−3 −2 −3
O1 O2 O1 O2 O1 O2
Fig. 5.21 Comparisons of Arousal (top row) and Valence (bottom row) values correspondent to the
two smells for the three datasets: whole subjects (column 1), male subjects (column 2) and female
subjects (column 3). Asterisks indicate statistical significant differences between Benzaldehyde
and Isovaleric acid (O1, O2)
Table 5.27 Confusion matrix of Pleasant smell Vs Unpleasant smell

using cvxEDA considering x, z and w feature set
Pleasant smell Unpleasant smell
K-NN Dataset (Benzaldehyde) (Isovaleric acid)
Pleasant smell All subjects 65.63 % 28.12 %
Males 56.25 % 31.25 %
Females 81.25 % 25 %
Unpleasant smell All subjects 34.37 % 71.88 %
Males 43.75 % 68.75 %
Females 18.75 % 75 %
The bold values represent the percentage of number of predictions/
classifications that were correct (true positives and true negatives)
5.7 Assessment of Mood States in Bipolar Patients

Using EDA
The results gathered from the multi-sensory applications suggested the possibility
to investigate psychiatric pathologies, which involve emotion disorders by means
of EDA analysis. In this research, we study the bipolar disorder. Bipolar patients
are characterized by a pathological unpredictable behavior, resulting in fluctuations
between states of depression and episodes of mania or hypomania. In the current
clinical practice, the psychiatric diagnosis is made through clinician-administered
rating scales and questionnaires, disregarding the potential contribution provided
by physiological signs. The aim of this research was to investigate how changes in
the autonomic nervous system activity can be correlated with clinical mood swings.
More specifically, a group of ten bipolar patients underwent an emotional elicitation
protocol to investigate the autonomic nervous system dynamics, through the EDA,
among different mood states. Physiological signals were analyzed by applying both
the CDA and the cvxEDA methods to decompose EDA into the tonic and phasic
components, from which several significant features were extracted to quantify the
sympathetic activation.
5.7.1 Patient Recruitment and Experimental Protocol
Ten patients affected by bipolar disorder I or II were selected for this study. None of
them had suicidal tendencies, delusions or hallucinations. Patients were admitted to
the psychiatric unit of the hospital and periodically screened through a psychiatric
interview. Before each acquisition a mood label among “euthymic”, “depressed”,
“maniac” and “mixed-state” was associated to each patient/acquisition. As a control
group, a group healthy subjects were enrolled and participate to the study. In
particular, ten healthy subjects (5 females, age ranged from 20 to 32), i.e. not
suffering from both cardiovascular and evident mental pathologies, was asked to
fill out the Patient Health QuestionnaireTM (PHQ). All participants showed score
lower than 5. Such a cut-off value was chosen in order to avoid the presence of
either middle or severe personality disorders [318].
An ad-hoc affective elicitation experimental was administered to both the healthy
and bipolar patients group. In particular, such an experimental protocol, graphically
shown in Fig. 5.22, was structured as follows:
• 5-min at rest with closed eyes;
• 5-min at rest with open eyes;
• 6-min slideshow of IAPS pictures with high arousal and negative valence;
• up to 4 min of pictures gathered from TAT.
Fig. 5.22 Block scheme of the experimental protocol

Fig. 5.23 Example of EDA signal and related components during euthymic state, extracted
through deconvolutive method of analysis. On the top panel, the black signal representing the raw
EDA signal along with the DRIVERtonic (red) are shown. On the lower panel, the DRIVERphasic is
shown. Rest phases lasted for the first 600 s. Afterwards, IAPS and TAT emotional stimulation is
performed
As described above, the protocol is split into two sessions: rest and emotional
elicitation. The latter session is divided, in turn, into two stages, both of which are
intended to elicit a variation of the ANS response. Specifically, IAPS pictures lasted
for 2 s presenting negative emotional contents (high arousal and negative valence).
The same IAPS pictures were presented to all patients and healthy subjects and
nobody was asked to score the elicited level of arousal and valence. The images
were chosen according to the following characteristics: arousal score > 6:7; valence
< 4:5. Afterwards, patients were invited to tell a story based on the input coming
from the TAT pictures. However, in order to avoid biased results related to the IAPS
and TAT sequential order, IAPS-TAT and TAT-IAPS session order was randomly
interchanged. The hypothesis of this study is that the ANS differentially reacts to
such emotional stimuli upon different pathological mood states. During the whole
duration of the protocol, the EDA signal was acquired using the BIOPAC MP150
system with a sampling frequency of 1000 Hz (Fig. 5.23). EDA sensors were placed
on the distal phalanx of the second and third finger of the non-dominant hand,
imposing a DC voltage of 0.5 V. The protocol was run for a follow-up period up
to 75 days. Patients repeated the protocol at each mood change, whereas healthy
subjects repeated the experiment twice within 2 weeks in order to investigate
Table 5.28 Clinical Acq.1 Acq.2

evaluations of the patients
Mood state Mood state
Pz01 Depressed Euthymic
Pz03 Mixed-state N.P.
Pz04 Mixed-state Euthymic
Pz05 Mixed-state N.P.
Pz06 Depressed N.P.
N.P. stands for Not Performed
Table 5.29 Features extracted from phasic and tonic components

Feature Description
MAX-Tonic Maximum value of the tonic driver curve
MAX-Phasic Maximum value of the phasic tonic curve
AUC-Tonic Area under the tonic driver curve over time
AUC-Phasic Area under the phasic driver curve over time
Mean-Tonic Mean value of the tonic driver component
Mean-Phasic Mean value of the phasic driver component
STD-Tonic Standard deviation of the tonic driver component
STD-Phasic Standard deviation of the phasic driver component
possible differences in the EDA pattern between repeated acquisitions during no

pathological mood states and swing. Of note, seven patients (i.e. Pz01, Pz02, Pz04,
Pz07, Pz08, Pz09, Pz10) were acquired twice, whereas Pz03, Pz05, and Pz06 carried
out a single acquisition. Details are shown in Table 5.28.
5.7.1.1 Statistical Analysis of Electrodermal Response to IAPS Stimuli

in Bipolar Patients
Skin conductance data related to the TAT sessions was excluded from the analysis
since patients’ voice could affect the EDA acquisition. Therefore we considered
only the IAPS stimulation.
Both cvxEDA and CDA model were applied to each EDA time series and several
features were extracted from both phasic and tonic components.
In Table 5.29 the whole set of features is reported along with a corresponding
description. Each feature was normalized by subtracting its correspondent value at
rest.
For each of the seven subjects (i.e. Pz01, Pz02, Pz04, Pz07, Pz08, Pz09 and Pz10)
who changed their mood state and performed the experiment twice, the feature-sets
extracted from the two different acquisitions were compared by using a Wilcoxon
test for paired data [319] (i.e., intra-subject statistical analysis).
Moreover, an inter-subject analysis was performed. For each features, all values
associated to the same mood label were grouped. The three different groups corre-
spondent to the three mood states (i.e. depression, mixed-state and euthymia) were
compared by means of a Kruskal-Wallis test to evaluate whether they statistically
belonged to the same population. In case of rejection of the null hypothesis a Mann-
Whitney post-hoc analysis [320] with Bonferroni adjustment was carried out.
Of note, the statistical inference analysis was performed by means of non-
parametric tests due to the non-gaussianity of the samples (p < 0:05 given by
Kolmogorov-Smirnov test with null hypothesis of Gaussian distributed samples).
5.7.1.2 Classification of Mood States in Bipolar Patients
The classification process aimed at performing the recognition of clinical mood

states such as DP, MX, EU. In this work a standard k-Nearest Neighbor (k-NN)
classifier was used. This choice is justified by ease of computation as well as the
capability of dealing with non-parametric data of this algorithm. Remarkably, this
classifier can be a viable solution to be implemented in wearable electronic devices.
k-NN predicts the class finding the k closest training points, and the new example
is assigned to the most common class amongst its k nearest neighbors. After the
training process, the performance of the classification task is commonly evaluated
using the confusion matrix. A more diagonal confusion matrix corresponds to a
higher degree of classification. The training phase is carried out on 80 % of the
feature dataset while the testing phase on the remaining 20 %. We performed 40-
fold cross-validation steps to obtain unbiased classification results. In this work,
classifications were performed considering the inter-subject variability. Inter-subject
classification considers features gathered from all the available acquisitions and
grouped considering the mood label exclusively.
5.7.2 Experimental Results
In this section, the experimental results performed on both groups of bipolar patients
and controls (i.e., healthy subjects) are shown in detail. Further statistical analyses
pointing out differences between phasic and tonic features, for each EDA model and
for each acquisition, as well as results on intra- and inter-subject evaluations follow
below (see Tables 5.30 and 5.31).
Results on Bipolar Group A summary of the clinical evaluations of the patients
recruited for this study, expressed as mood labels, is shown in Table 5.28.
Concerning the CDA results, for each acquisition, we found significant dif-
ferences (p < 0:03) for all of the considered EDA features but the STD-Tonic.
Table 5.30 Results from the bipolar patients dataset expressed as statistical significance
for each EDA feature extracted from the CDA model outputs
IAPS Pz01 Pz02 Pz04 Pz07 Pz08 Pz09 Pz10
MAX-Tonic < 106 > 0:05 > 0:05 < 106 < 106 < 106 < 106
MAX-Phasic < 104 < 106 < 106 < 106 < 106 < 106 < 106
AUC-Tonic < 106 > 0:05 > 0:05 < 106 < 106 < 106 < 106
AUC-Phasic < 104 < 106 < 106 < 106 < 106 < 0:05 < 106
Mean-Tonic < 105 > 0:05 > 0:05 < 106 < 106 < 106 < 106
Mean-Phasic < 104 < 106 < 106 < 106 < 106 < 0:05 < 106
STD-Tonic < 0:05 > 0:05 > 0:05 > 0:05 > 0:05 < 0:005 < 104
STD-Phasic < 104 < 106 < 104 < 106 < 106 < 106 < 106
p-values are from the Wilcoxon test
Samples are estimated during IAPS elicitation sessions of the two acquisition/mood states
The bold values represent the significant statistical values
Table 5.31 Results from the bipolar patients dataset expressed as statistical significance
for each EDA feature extracted from the CvxEDA model outputs
IAPS Pz01 Pz02 Pz04 Pz07 Pz08 Pz09 Pz10
MAX-Tonic > 0:05 > 0:05 > 0:05 > 0:05 > 0:05 > 0:05 > 0:05
MAX-Phasic < 106 < 106 < 106 < 106 < 106 < 106 < 106
AUC-Tonic > 0:05 > 0:05 > 0:05 > 0:05 > 0:05 > 0:05 > 0:05
AUC-Phasic < 106 < 106 < 106 < 106 < 106 < 103 < 106
Mean-Tonic > 0:05 > 0:05 > 0:05 > 0:05 > 0:05 > 0:05 > 0:05
Mean-Phasic < 106 < 106 < 106 < 106 < 106 < 106 < 106
STD-Tonic < 103 < 103 < 106 < 103 < 103 < 0:01 < 103
STD-Phasic < 106 < 106 < 106 < 106 < 106 < 106 < 106
p-values are from the Wilcoxon test
Samples are estimated during IAPS elicitation sessions of the two acquisition/mood states
The bold values represent the significant statistical values
CvxEDA algorithm showed similar results about phasic components, instead the
tonic features were not significant except for the STD-Tonic (Tables 5.31 and 5.32).
Results Using CDA Model Wilcoxon test for paired data was applied on patients
with two acquisitions, i.e Pz01, Pz02, Pz04, Pz07, Pz08, Pz09 and Pz10. Statistical
analysis results show that all the phasic features resulted to be statistically different
for all subjects. Patients Pz02, Pz04 showed a non-significant tonic features set
between the two acquisitions. More in detail, patients Pz01, Pz07, Pz08, Pz09
and Pz10 exhibited significant increase in the mean value, in the area under the
curve and in the maximum value of both DRIVERphasic and DRIVERtonic components
during second acquisition (see an example in Fig. 5.24). Pz02 showed no statistical
difference in tonic features, but an increasing significant trend of the phasic features
was found. As all of five patients clinically improved (i.e. change into an euthymic
state) their status, this results could be due to an increased sympathetic activity
during the emotional stimulation session [227]. On the contrary, Pz04 showed a
Table 5.32 Specification of increasing or decreasing trends of cvxEDA phasic components

during clinical mood swings
Pz01 Pz02 Pz04 Pz07 Pz08 Pz09 Pz10
Depressed Depressed Mixed-st Depressed Mixed-st Mixed-st Depressed
Euthymic Euthymic Euthymic Euthymic Euthymic Euthymic Euthymic
MAX-Phasic " " # " " " "
AUC-Phasic " " # " " " "
Mean-Phasic " " # " " " "
Up-arrow and down-arrow intend an activity increase and decrease between the two acquisi-
tions, respectively
(a) AUC PHASIC DRIVER
1
acquisitions
40 50 60 70 80 90 100 Ranks110
(b) AUC TONIC DRIVER
1
acquisitions
5 10 15 20 25 Ranks 30
Fig. 5.24 Pz01’s statistical analysis for IAPS elicitation. Results of Pz01’s CDA-AUC of
DRIVERphasic (a) and DRIVERtonic (b) features
significant decrease for all phasic features in the second acquisition as compared to
the first one, whereas tonic features were not statistically different (see an example
in Fig. 5.25). Yet this result can be interpreted as a reduction of sympathetic activity
when moving from a mixed state, where hypomanic symptoms could be present,
to an euthymic condition [227]. The standard deviation of both DRIVERtonic and
DRIVERphasic components showed similar trend between the two acquisitions for all
of the seven patients having two observations. In particular, STD-Tonic and STD-
Phasic decreased in the second acquisition, i.e. euthymic state.
Furthermore, an inter-subject statistical analysis was performed including also
the patients with one acquisition only. Data were not considered as coming from
AUC PHASIC DRIVER

(a)
1
acquisitions
30 40 50 60 70 80 90 100Ranks110
AUC TONIC DRIVER
(b)
1
acquisitions
10 12 14 16 18 20 22 Ranks 24
Fig. 5.25 Pz04’s statistical analysis for IAPS elicitation. Results of Pz04’s CDA-AUC of
DRIVERphasic (a) and DRIVERtonic (b) features
a specific subject but grouped following clinical classification. A Kruskal-Wallis

test was carried out among acquisitions classified as depressed, mixed-state and
euthymic. The mean value and AUC of the DRIVERphasic signal significantly
discriminated the three mood states (p < 106 ). In particular, this two features
exhibited the same trend, i.e. they increase from depression to euthymia through
mixed-state (see Fig. 5.26a).
The maximum value of DRIVERphasic was able to distinguish the depressed mood
states from the group mixed-state plus euthymic state (p < 106 ). Instead, mixed-
state and the euthymic state did not show a significant difference (p > 0:05).
Concerning features extracted from DRIVERtonic (see Fig. 5.26b), the Kruskal-
Wallis test showed significant differences among the three different mood states
(p < 106 ), despite the fact that the depression and mixed-state group and the
depression and euthymic group did not show significant difference (p > 0:8).
Results Using cvxEDA Model The cvxEDA tonic features did not show any
significant changes on all of the subjects in the transition from the depressed or
mixed-state to the euthymic one, except for the standard deviation of the tonic signal
that increased during the euthymic state. Considering the phasic features, using the
cvxEDA we found the same trend and statistically significant results as described
for the CDA model (Pz01 and Pz04 examples of phasic-AUC feature are shown in
Figs. 5.27 and 5.28).
Even for the cvxEDA model was carried out an inter-subject statistical analysis.
The tonic features did not show significant differences among the three mood states
(a) AUC PHASIC DRIVER
Depressed
Mixed−State
Euthymic
350 400 450 500 550 600 650

Ranks
(b) AUC TONIC DRIVER
Depressed
Mixed−State
Euthymic
80 90 100 110 120 130 140 150 160

Ranks
Fig. 5.26 IAPS stimulation: inter-subject statistical analysis. CDA-AUC of DRIVERphasic (a) and
DRIVERtonic (b) features
(see example in Fig. 5.29b). In the phasic AUC features results confirmed the CDA
considerations, but with an increased level of the statistical significance (i.e., a
lower p-value), especially in the comparisons between depressed or mixed-state and
euthymic state (see Fig. 5.29a). A different contribution of the cvxEDA model was
shown for the inter-subject statistical analysis of the maximum value of the phasic
component. In this case, the features was able to distinguish the three mood states
and not only the depressed (p < 103 ) from the group mixed-state plus euthymic
state.
Results on Healthy Controls We performed statistical analyses based on the
Wilcoxon test for paired samples to investigate whether differences on the EDA
feature patterns of healthy subjects are statistically significant between multiple
affective elicitation protocols over time. Likewise the analysis performed on the
bipolar patients group, features reported in Table 5.35 were extracted form both
the phasic and tonic series. We report that the inter-subject statistical analysis
independently performed considering data from IAPS and TAT sessions showed
no statistically significant differences between the two acquisitions on each of the
considered EDA features (p > 0:05).
Classification Results In order to verify whether the proposed methodologies were
able to recognize changes in the ANS dynamical patterns associated to different
mood states, the capability of the k-NN algorithm for solving the 3-class inter-
subject pattern recognition problem was also tested. The extracted features were
grouped in three sets. Specifically, we defined the feature set ˛ as the set extracted
from driver tonic exclusively; feature set ˇ as the set extracted from driver phasic
(a)
1
acquisitions
150 200 250 300 350 400
(b)
1
acquisitions
260 270 280 290 300 310 320
Fig. 5.27 Pz01’s statistical analysis for IAPS elicitation. Results of Pz01’s CvxEDA-AUC of
phasic (a) and tonic (b) features
(a)
1
acquisitions
200 220 240 260 280 300 320 340
(b)
1
acquisitions
260 265 270 275 280 285 290 295 300
Fig. 5.28 Pz04’s statistical analysis for IAPS elicitation. Results of Pz04’s CvxEDA-AUC phasic
(a) and tonic (b) features
5.8 Changing Source Oscillations of Skin Admittance: A Study in the. . . 103
(a)
Depressed
Mixed−State
Euthymic
1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 2600
(b)
Depressed
Mixed−State
Euthymic
1980 2000 2020 2040 2060 2080 2100 2120
Fig. 5.29 IAPS stimulation: inter-subject statistical analysis. CvxEDA-AUC of phasic (a) and
tonic (b) features
exclusively; and feature set as the set obtained as union of ˛ and ˇ sets. The
classification procedure was performed using all the three feature sets, in order to
assess the contribution of the tonic and phasic components of the EDA. Results are
expressed in form of confusion matrices. Results achieved using the feature set
coming from the CDA analysis are very satisfactory and reported in Tables 5.33
and 5.34. Indeed, confirming the evidences on the inter-subject statistical analysis,
when using both tonic and phasic information the system is able to achieve
accuracies always greater than 80 %.
The cvxEDA model confirmed an accuracy over the 80 % with the whole dataset,
but increased the accuracy of the pattern recognition with both the phasic and tonic
dataset taken separately.
5.8 Changing Source Oscillations of Skin Admittance:

A Study in the Frequency Domain with Application
on Emotion Recognition
There are two main methods for measuring EDA: endosomatic (internal electrical
source) and exosomatic (external electrical source). Even though the exosomatic
approach is the most widely used, differences between alternating current (AC)
and direct current (DC) methods and their implication in the emotional assessment
field have not yet been deeply investigated. This section aims at investigating how
Table 5.33 Confusion K-NN DP MX EU

matrix of DP vs MX vs EU
using CDA features from all DP ˛ 49.05 ± 10.78 10.77 ± 9.21 10.00 ± 10.28
of the samples regardless any ˇ 60.12 ± 4.93 10.37 ± 2.51 17.66 ± 3.98
specific patient 80.00 ± 4.96 8.50 ± 2.17 7.57 ± 2.39
MX ˛ 31.43 ± 7.17 67.69 ± 9.28 20.38 ± 7.03
ˇ 18.49 ± 4.57 75.33 ± 4.97 17.38 ± 4.21
9.77 ± 2.91 83.27 ± 2.13 8.22 ± 3.49
EU ˛ 19.52 ± 7.26 21.54 ± 9.100 69.62 ± 9.32
ˇ 21.39 ± 4.46 14.30 ± 3.47 64.95 ± 5.47
10.23 ± 3.28 8.22 ± 1.69 84.20 ± 3.92
The bold values represent the percentage of number of predic-
tions/classifications that were correct (true positives and true
negatives)
Table 5.34 Confusion K-NN DP MX EU

matrix of DP vs MX vs EU
using cvxEDA features from DP ˛ 58.70 ± 2.53 18.02 ± 2.46 23.29 ± 2.63
all of the samples regardless ˇ 74.18 ± 2.98 26.18 ± 2.36 15.14 ± 2.58
any specific patient 80.05 ± 2.98 9.85 ± 1.96 13.57 ± 2.18
MX ˛ 17.68 ± 2.15 67.74 ± 2.71 14.29 ± 2.20
ˇ 10.34 ± 2.44 62.30 ± 2.98 11.78 ± 1.75
9.01± 2.05 80.30 ± 2.28 5.81 ± 1.92
EU ˛ 23.62 ± 2.42 14.23 ± 2.18 62.41 ± 2.45
ˇ 15.47 ± 2.36 11.51 ± 1.72 73.08 ± 2.37
10.93 ± 2.40 9.85 ± 1.76 80.62 ± 2.51
predictions/classifications that were correct (true positives
and true negatives)
the admittance contribution of EDA, studied at different frequency sources, affects

the EDA statistical power in inferring on the subject’s arousing level (neutral or
aroused). To this extent, 40 healthy subjects underwent visual affective elicitations,
including neutral and arousing levels, while EDA was gathered through DC and AC
sources from 0 to 1 kHz.
5.8.1 Experimental Protocol
Forty healthy subjects were enrolled in the experiment, aged 26˙ 4 (18 females).
All subjects gave written informed consent before taking part in the study, which
was approved by the local Ethics Committee. The experiment was designed as
following:
• initial resting phase of 1 min;

• maximal expiration task phase of about 1 min;
• affective visual stimulation phase of 2 min;
• final resting phase of 1 min;
(The two elicitation phases will be described in detail in the next sub-sections.)
Subjects were comfortably seated in an acoustically insulated room in front of a
computer screen while their EDA was recorded using the presented acquisition
system.
Note that the group of 40 healthy subjects was split into four subgroups, each
of which comprised of 10 subjects. Each subgroup were acquired with a different
exosomatic method such as DC (group 1), AC with a frequency of 10 Hz (group 2),
AC with a frequency of 100 Hz (group 3) and AC with a frequency of 1 kHz
(group 4).
Maximal Expiration Task In this session of the experiment, all of the 40 subjects
performed a forced maximal expiration task [89], in which they were asked to
breathe out with the maximum possible intensity in order to trigger the SNS-
mediated expiration reflex.
After the initial resting state session, the subjects breath normally and rest in front
of the computer monitor for about 20 s, then they had to perform a deep expiration
twice with an inter stimulus interval of about 20 s, after a neutral visual input on the
screen.
The use of the forced expiration task is justified by the need of having a
stimulus whose EDA response was as reliable and objective as possible. In fact,
previous studies have demonstrated that this stimulation is a reliable way to evoke
phasic responses unaffected by emotional change with better reproducibility, less
habituation, and more stable waveform patterns than other experimental paradigms
(including electrical) [89]. In this way, the presence of at least one phasic response
after each stimulus was ascertained. Therefore, we could investigate whether the
cvxEDA algorithm was able to identify each phasic response for each acquisition
method.
Affective Visual Stimulation In the second elicitation session, each group of 10
participants was stimulated by projecting on a screen images selected from the
official IAPS (International Affective Picture System) database [90]. IAPS dataset
is a collection of images ranked in terms of arousal (i.e. intensity of perception) and
valence (pleasantness of perception). This protocol session is designed to assess
the pattern recognition system ability on each data group (i.e., of each method) to
correctly classify stimulations with different arousal content and provide meaningful
information about SNS activation.
The slideshow timeline consists in 3 neutral images, 6 aroused images and 3
other neutral images. Each image was shown for 10 s.
5.8.2 EDA Analysis and Classification Procedure
For each dataset, the cvxEDA model was applied to each time series. Concerning
the IRF parameters considered for this study, values of 1 D 0:7 s, 0 D 0:7 s, ˛ D
0:4 and D 0:01 were employed throughout this analysis, according to previous
exploratory tests on separate data.
In the respiratory stimulation dataset, the presence of an estimated burst of
SMNA activity was verified in each 5-s time window following a stimulus onset,
in order to prove the model’s ability to correctly detect partially overlapped phasic
responses.
As summarized in Table 5.35, we segmented each signal in correspondence to
each IAPS image time window and we extracted several features from both the
tonic and phasic component:
5.8.2.1 Summary of the Classification Procedure
The feature set, extracted from each single IAPS image, was used as input of a
pattern recognition algorithm in order to classify the two arousal levels, according
to the IAPS rates. The supervised classification of the feature set was implemented
following a Leave-One-Subject-Out procedure (LOSO) applied to a KNN-based
classifier. For each of the N iterations (where N is the total number of participants)
the whole dataset was split into a training set including .N 1/ subjects and
a test set including the cvxEDA feature values of the remaining subject Nth.
Moreover for each iteration of the LOSO scheme, a feature selection procedure was
performed in order to identify the combination of parameters that resulted in the
highest recognition accuracy within the training set examples. Each selected feature
constituted a single dimension of the feature space. The LOSO pattern recognition
procedure is illustrated in Fig. 5.30.
Table 5.35 List of features extracted from EDA phasic and tonic compo-
nents
Feature Description
Npeak Number of significant SMNA peaks wrw
AUC Area under curve of reconstructed phasic signal wrw (S s)
MeanTonic Mean value of the tonic component
within each image time window (S)
wrw within response window (i.e., 5 s after stimulus onset)
Decomposition
N-1 Selection
classification
Fig. 5.30 Overall block scheme of the proposed valence recognition system. The EDA is
processed in order to extract the phasic and tonic components using the cvxEDA algorithm.
According to the protocol timeline, several features are extracted. The KNN classifier is engaged
to perform the pattern recognition by adopting a leave-one-subject-out procedure
5.8.3 Classification Results
As we expected from the specifics of the cvxEDA model, all EDA data (Fig. 5.31a)
were decomposed into two signals, a sparse component p and a smooth component
t, that we interpret as the activity of the sudomotor nerve (Fig. 5.31b) and the tonic
level (Fig. 5.31c).
Maximal Expiration Task Results We performed both a visual and a statistical
inspection of time series to verify whether the effectiveness of the experimental
protocol in eliciting phasic responses was confirmed for all different kinds of
acquisition method (DC and AC).
After the application of the cvxEDA model, we considered a time windows of
5 s after the onset of each expiration task, and we looked for peaks of the SMNA
signal (in fact, the phasic response is defined as the part of the signal arises within a
predefined response window of 1–5 s [15, 16]) Of note, due to the stimulus intervals
of about 20 s no overlap between consecutive responses occurred.
Results of an inter-subject analysis showed that cvxEDA was able to correctly
detect the corresponding phasic peak response over 97:5 % of the respiratory stimuli.
Moreover, a visual inspection of the small percentage of cases that were not
correctly identified revealed a very low signal-to-noise ratio of that segments of
signal.
Automatic Arousal Recognition Results Results of the arousal-level-classification-
procedure on the four datasets, namely, DC, AC 10 Hz, AC 100 Hz, AC 1 kHz, are
shown in Tables 5.36, 5.37, 5.38, and 5.39. The recognition accuracy is reported in
the form of a confusion matrix. An element rij of the confusion matrix indicates a
percentage of mismatches, i.e. how many times a pattern belonging to class i was
erroneously classified as belonging to class j. Terms rij on the main diagonal of the
confusion matrix correspond to correct classifications.
(a) EDA raw data

2.5
[a.u.]
1.5
0 50 100 150 200 250 300
time [s]
(b) SMNA
30
20
[a.u.]
10
0
0 50 100 150 200 250 300
time [s]
(c) Tonic
2.4
2.2
[a.u.]
1.8
1.6
0 50 100 150 200 250 300
time [s]
Fig. 5.31 Application of the cvxEDA decomposition procedure to the EDA signal recorded for a
representative subject. (a) Raw EDA signal, Z-score normalized. (b) Estimated sparse phasic driver
component p. (c) Estimated slow tonic component t
Table 5.36 Confusion K-NN (DC) Neutral Arousal

matrix of Neutral Vs Arousal
images using cvxEDA feature Neutral 63.3333 % 35.00 %
set extracted with DC source Arousal 36.67 % 65.00 %
age of number of predictions/classifi-
and true negatives)
Both DC and AC measures did not show very high average recognition accuracy.
However, it is worthwhile noting that using 100 Hz of frequency current source,
we obtain an average accuracy significantly higher than in the other cases. More
specifically, using DC, 10 Hz and 1 kHz, the average accuracy was in the range
of 62.5–63.34 %, whereas at 100 Hz the pattern recognition system showed an
accuracy of 71.67 %.
Table 5.37 Confusion K-NN (AC 10 Hz) Neutral Arousal

set extracted with AC source Arousal 35.00 % 61.67 %
at 10 hz The bold values represent the percentage
of number of predictions/classifications that
were correct (true positives and true nega-
tives)
Table 5.38 Confusion K-NN (AC 100 Hz) Neutral Arousal

set extracted with AC source Arousal 31.67 % 75.00 %
at 100 hz The bold values represent the percentage
of number of predictions/classifications that
were correct (true positives and true nega-
tives)
Table 5.39 Confusion K-NN (AC 1 kz) Neutral Arousal

images using cvxEDA feature Neutral 63:33 % 38:33 %
set extracted with AC source Arousal 36:67 % 61:67 %
at 1 khz The bold values represent the percentage
negatives)
Chapter 6
Conclusions
This book has unveiled the strong relationship between Electrodermal Activity
(EDA) signal and autonomic nervous system (ANS) dynamics, and how EDA
could be source of reliable and effective markers for the characterization of the
physiological response to different emotional stimuli and for the automatic affective
and mood state recognition.
In the literature, many studies have demonstrated the link between EDA and
ANS. It is worthwhile mentioning that until the early 1990s, most of the analyses
of the EDA relied, almost exclusively, on heuristic methods, such as the visual
inspection. In the last two decades, a model-based approach has emerged and several
mathematical models have been developed in order to automatize the decomposition
and the processing of the EDA signals as well as the feature extraction stage. In fact,
many studies have examined automatic ways to count spontaneous SCRs, to extract
amplitude or other measures of a single causal SCR, and to deal with motion artifacts
and superposition on the SCRs (there are also publicly- available toolboxes for these
tasks) [16, 66]. Such automatic methods have brought the opportunity to estimate
the unobservable processes (e.g. the sudomotor nerve activity, SMNA) underlying
the EDA phenomena, along with the relationship between SMNA and the sweat
diffusion process. Nevertheless, they still relied, in part, on the use of post- and
pre-processing stages and ad-hoc solutions. Using the model-based approach, the
extracted features have demonstrated to be reliable enough while inferring on the
central nervous system. For example, a recent study proposed a feature that allows
to quantify the sympathetic activity from EDA [85]. This has overcome issues that
are still present in other physiological signals, such as HRV.
Throughout the book, conventional methodologies of EDA processing and
models have been described, also aiming to perform a fair comparison with our
recently proposed EDA processing through convex optimization approach, whose
use has been indeed emphasized in some chapters. Such a modeling approach, called
cvxEDA, was based on maximum a posteriori probability, convex optimization,
and sparsity. The model describes the recorded skin conductance signal as the sum

112 6 Conclusions
of three terms: the phasic component, the tonic component, and an additive white
Gaussian noise term incorporating model prediction errors as well as measurement
errors and artifacts. The new algorithm models the IRF (2.9) as an ARMA model
(i.e. an IIR filter) instead of a MA model (i.e. a FIR filter). This allows a much
more compact representation of the IRF by means of two tridiagonal matrices
instead of a banded matrix, thus increasing the accuracy and significantly reducing
the computational cost. In fact, the sparsity and structure of the problem (2.21)
can be effectively exploited by the state-of-the-art sparse-QP solvers. The main
difference between this model other methods in the literature lies in the presence
and definition of the prior probabilities for the phasic and tonic signals. Positiveness
and burstiness of the sudomotor nerve activity driving the phasic component
is modelled through a first order description of spike trains, i.e. assuming a
Poisson distribution approximated by an exponential distribution. This form of
the prior probability translates into a non-negative inequality constraint and an l1 -
norm regularizer in the final optimization problem. Although one could impose
a stronger regularization—e.g. l0 -“norm” [82]—on the phasic driver, this would
render the problem non-convex, i.e. computationally more demanding, and would
significantly deviate from the physiological explanation in terms of Poisson spike
trains. Physiologically-plausible temporal scale and smoothness of the tonic input
signal are achieved by means of an adequate choice of the spacing between the
knots of the spline and through a Gaussian prior on the values at the knots, which
ultimately translates into an l2 regularization of the spline’s coefficients in the
optimization problem. Thanks to the ARMA observation model and to this choice
of priors, physiologically sound constraints on the signals can be imposed to be
estimated and yet be able to obtain the globally optimal solution by solving a
standard quadratic-programming problem. The proposed cvxEDA model shares
some major limitations with most state-of-the-art algorithms, mainly by relying on
the strong assumptions of linearity and time-invariance of the system. In reality,
physiological systems—especially those involving neural dynamics—are likely
to show nonlinear and complex dynamics. Furthermore, such a dynamics and
its statistical properties can be different among subjects and further depend on
environmental and experimental conditions. Within the proposed EDA modeling
framework, inter- and intra-subject variability can be accounted for by choosing
a customized IRF function for each subject/condition. This problem was partially
addressed in our experimental analysis by performing an outer optimization step to
tune the slow time constant of the IRF for each specific subject.
The new algorithm was evaluated in three ways to test its robustness to noise,
its ability to separate and identify each phasic response (even when they overlapped
because of short ISIs) and its capability of properly describing the activity of the
autonomic nervous system in response to strong affective stimulation. The results
of the three analyses confirmed the proprieties of the model. On a synthetic dataset,
the algorithm proved to be robust to different levels of noise. When applied to real
data from a forced maximal expiration protocol, the algorithm demonstrated strong
ability to reliably detect phasic responses to eliciting stimuli, also overcoming the
problem of overlapping SCRs encountered in experimental paradigms involving
6 Conclusions 113
short ISIs. In the affective stimulation paradigm, the mean tonic level estimated by
the model was significantly different in arousal and neutral sessions. Analyzing the
phasic response, we found a consistent statistical relationship between the arousal
levels and the peak amplitude of the estimated phasic driver, thus confirming the
model’s predictive validity. These results were compared to those obtained using
Ledalab implementation of the CDA [1], a method that performs a deterministic
inversion of the peripheral model. The trends found using the CDA confirmed those
obtained from cvxEDA model. However, CDA only found statistically significant
differences between the lightest and the strongest levels of arousal while cvxEDA
model allowed a finer discrimination.
Because it can be implemented in few lines of code and does not depend
on external libraries (except for a conventional QP solver), the new algorithm
has a wide applicability and can be readily integrated in existing open-source
psychophysiological modeling software. Given all these features and the low
computational cost of the proposed algorithm, the cvxEDA model can be employed
in further affective computing applications.
In order to evoke changes in the human affective states in a laboratory setting, we
assembled sets of pictures, sounds, caresses and smells chosen to elicit a range of
aroused, positive, neutral or negative emotions. Valence and arousal were identified
as the principal dimensions of affective response to the environment. In this case,
valence is defined as the degree to which one has favorable feelings towards a
situation, while arousal is defined as the degree to which one feels excited in the
situation. EDA and its ability to characterize the emotion reaction in a multi-sensory
scenarios is studied in four different configurations.
Affective Visual Scenario Concerning the affective visual stimulation, in addition
to the statistical analysis among different levels of arousal, described above, we
have shown a multivariate pattern recognition analysis considering both tonic and
phasic features. More specifically, the set of features were extracted using cvxEDA
and CDA and used as input to a classifier to automatically recognize four arousal
classes and two valence levels.
EDA features allowed achieving a greater sensitivity to changes in the arousal
levels than in the valence levels of the stimulus. Using the cvxEDA approach, results
were very satisfactory and all the four classes of arousal and the two classes of
valence could be discriminated with an acceptable error, i.e. we obtained over 71 %
of successful recognition for the arousal problem and over 68 % for the valence
one. The CDA showed similar percentages in case of valence classification, but its
performance remarkably dropped in the arousal recognition problem. Therefore,
looking at these results, we could assume that the contribution provided by the
convex optimization approach is essential to attain much better results.
Affective Auditory Scenario Concerning the auditory stimulation, we applied the
cvxEDA approach used for the visual scenario to automatically recognize emotions,
as elicited by affective sounds, in young healthy subjects. Even in this case emotions
were expressed in terms of arousal and valence levels according to the circumplex
114 6 Conclusions
model of affect. The experimental protocol consisted in three levels of arousal

stimuli, each of which containing pleasant and unpleasant sounds.
The statistical comparison among the three arousal levels and the two valence
levels showed no significant differences on the phasic feature set. On the contrary,
the mean tonic value, computed from the cvxEDA outputs only, was significantly
different among the three arousal levels and the two valence levels. One possible
explanation can be related to the long duration of each experimental session
and the possible consequent loss of concentration of the subjects. On the other
hand, such long stimuli were able to induce a significant change in the general
psycho-physiological state of the subjects that was highly correlated with the tonic
component, as highlighted by the tonic features.
The classification procedure was applied to discriminate the two valence levels
and the three arousal levels. Results achieved using the cvxEDA model showed
a recognition accuracy of 77.33 % on the arousal dimension, and of 84 % on
the pleasantness dimension. Likewise the statistical analysis, the feature selection
in the classification procedure suggested that ANS measures related to the tonic
component were the most effective in discriminating and recognizing emotional
states induced by affective sounds. This finding confirms the relevant contribution
of the cvxEDA model, which determines an increase of the accuracy of more than
30 % compared with the CDA.
To our knowledge, this study shows for the first time the use of sympathetic
nervous system measures, such as the EDA, to recognize emotional arousal states
induced by affective sounds. It is also worthwhile noting that we were able to discern
arousing stimuli by using features extracted from the EDA series exclusively. This
achievement along with previous results (i.e., visual stimulation) are so promising
to open new avenues in the field of affective computing, suggesting that emotion
recognition is possible using data coming from EDA only. Of note, emotion
recognition through EDA only was already performed using visual stimuli [57].
Moreover, outcomes of this research might have an impact in clinical research fields
such as psychology or neurology. The use of visual emotional elicitations, in fact,
could not be applied to patients with visual impairment, or patients with disorder of
consciousness. Knowledge on sympathetic dynamics, as estimated through cvxEDA
models, in healthy subjects might provide useful biomarkers to support clinical
decisions.
Affective Touch Scenario In this scenario, we proposed the cvxEDA algorithm as
optimal method to perform EDA processing through wearable monitoring systems.
In particular, we used this approach to effectively discern caressing force and
velocity levels as elicited through a wearable haptic system [247]. EDA recordings
were gathered from 32 healthy subjects (16 females) undergoing affective haptic
elicitation through a wearable system able to convey caress-like stimuli by means
of two motors, which stretch a strip of fabric. Six kinds of stimuli were randomly
administered in time, being comprised of combinations of three velocities and two
forces levels.
6 Conclusions 115
Performance of the new cvxEDA technique was compared with the ones
obtained through the CDA model [15]. Comparisons were performed on statistical
significance in discerning affective stimuli along the force and velocity dimensions,
the time of execution of the algorithm, and memory usage.
Concerning the study of the statistical power of the EDA features, experimental
results demonstrated that, along the caressing force dimension, the cvxEDA and
CDA models have the same discriminant power. Only the diffTonic feature, in
fact, was significantly different between the two caressing force levels, with higher
caressing force associated with higher feature values. This means that the higher the
intensity of caressing, the higher the tonic level of EDA after such a cutaneous
stimulus. This is a reasonable behavior, being in line with typical physiological
dynamics associated to EDA [15]. However, it is worthwhile noting that the cvxEDA
provided tonic features with more discriminant power (i.e., lower p-values) than the
CDA approach.
Along the caressing velocity dimension, experimental results demonstrated that
cvxEDA modeling outperforms classical CDA approach. Features from the sparse
phasic components of EDA, in fact, were able to discern all of the differences
between caressing velocity levels. Phasic components estimated from CDA, instead,
were not able to discern between V2 vs. V3, and were always associated with higher
p-values than the cvxEDA ones (see Tables 5.14 and 5.15). Importantly, increasing
monotonic trends among caressing velocities were associated to cvxEDA-related
phasic features. Therefore, it is possible to conclude that cvxEDA modeling
approach provides feature values able to automatically assess caressing stimuli in
a force-velocity space.
We also demonstrated that the cvxEDA approach is particularly suitable for
implementations in embedded computing systems. Computational performance
analysis, in fact, demonstrated that the execution time of the cvxEDA algorithm
linearly increases with the length of the acquisition, whereas processing time of
CDA model tends to grow superlinearly (see Fig. 5.18). This is reasonably due to
the CDA optimization stage [15]. Moreover, taking advantage of the sparse nature of
its phasic components, cvxEDA-derived outputs needed significantly lower storage
values than the CDA model, thus being more suitable for the implementation in
wearable monitoring systems than CDA. Furthermore, it is worthwhile noting that
the cvxEDA approach needs to solve a convex optimization problem, thus always
guaranteeing to find the globally optimal solution. Moreover, it is worth to notice
that the degree of sparsity of the cvxEDA-based phasic components depends on the
number of peaks, i.e., number of stimulus responses, occurring in a given recording.
Therefore, differences in memory storage between the CDA and cvxEDA models
could be minimized even in other experimental protocols involving high frequency
stimuli.
Concerning the pattern recognition analysis, the two models showed opposite
performance with respect to the velocity and the force classification. The CDA
features showed greater accuracy than cvxEDA in the velocity classification,
whereas the cvxEDA parameters performed better accuracy in the classification of
the force levels. We can find an explanation of that, looking at the affective analysis
116 6 Conclusions
of caressing stimuli used in this study. We demonstrated how caressing force and
velocity levels relate to perceived emotional arousal and pleasantness levels of
emotions [157]. Specifically, subjects’ self-ratings revealed that caresses performed
at low force and low velocity are perceived as more pleasant and less arousing than
others [157, 211, 212]. However, while concerning the forces, we could make a
clear distinction between what force level was more pleasant and aroused, regarding
the velocities, the distinction was not so clear cut. As a consequence of these
considerations, we performed a new classification in terms of arousal and valence
scores of the participants. We identified two groups of arousal (i.e., aroused and
neutral stimuli) and valence scores as input of a SVM classifier. CvxEDA showed
a discrimination accuracy over the 75 % (10 % more than CDA) for both the
recognition problems, confirming the good ability in the emotion identification,
defined in terms of valence and arousal.
Findings of this study can be profitably exploited in the field of affective
haptics or, more in general, wearable haptic devices [321]. These systems, in
fact, require processing algorithms with low- computational cost and low-memory
consumption, in order to effectively augment communication, interaction, and
cooperation between human and robots.
Affective Olfactory Scenario The last experimental application concerns the
affective olfactory stimulation. In this study, an automatic valence recognition of
affective olfactory stimuli was performed. Specifically, we studied ANS dynamics
through the analysis of EDA in 32 healthy subjects (16 males). The experimental
protocol foresaw five different smells: Vanillin, Benzaldehyde, N-butanol, Isovaleric
acid, and Butyric acid. All the participants to the experiment had the same olfactory
threshold. Subjective ratings on SAM questionnaires confirmed that the Vanillin and
the Benzaldehyde were perceived as more pleasant, whereas the isovaleric acid and
the Butyric acid were assessed as unpleasant smells. No significant differences were
found among the arousal scores of the five smells.
EDA statistical analysis was carried out firstly studying the general physiological
state analyzing the tonic component (not directly related to the stimulus) [23] during
the resting states before and after the olfactory stimulus. Secondly, we studied
the stimulus-related responses in the time response window correspondent to the
presentation of each smell. Both analyses carried out on features extracted from the
CDA and cvxEDA models did not show any significant results.
Considering the challenging task of discerning valence levels from EDA signal
only, experimental results were quite satisfactory. We obtained an overall recogni-
tion accuracy of 68.75 % using a LOSO procedure on PWR-KNN classifier for both
the cvxEDA and the CDA models. The quite satisfactory classification accuracy
were obtained grouping the two pleasant smells and the two unpleasant smells as
assessed by the subjects.
Although the not statistical significant results, the classification problem suggests
remarkable electrodermal variations occurring during affective olfactory elicitation.
Accordingly, previous studies highlighted strong correlations between the effects of
odorous stimuli on EDA and valence, intensity and familiarity of the smells.
6 Conclusions 117
We also grouped olfactory data by gender. The experimental protocol was

reduced to two odorants among the initial five according to the SAM results
selecting the intermediate arousal levels: benzaldehyde and isovaleric acid.
Subjective ratings on SAM questionnaires confirmed that the benzaldehyde was
perceived as pleasant, whereas the isovaleric acid was assessed as unpleasant smell,
without significant difference between women and men.
The EDA of each participant was analyzed by means of the cvxEDA algorithm in
order to decompose the SC signal into its phasic and tonic components and extract
relevant features. Considering the challenging task of discerning valence recognition
problem, an overall recognition accuracy of 68.76 % was obtained using the LOSO
procedure on KNN classifier. A more satisfactory average accuracy (over 78 %)
was obtained only using feature form female set that show a much higher accuracy
then men (62.5 %). These results suggest a significant gender difference on EDA
dynamics in response to hedonic olfactory stimulation.
Accordingly, previous studies highlighted strong correlations between the effects
of odorous stimuli on EDA and valence, intensity and familiarity to the smells.
Moreover, prior art on fMRI studies highlighted different effects of hedonic odors
on gender, mainly related to the higher perception power of pleasantness associated
to the stimulus in women [316]. Our results also suggest a gender-specific change in
EDA as a response to pleasant smell stimuli. In fact, by taking into account gender
information, the accuracy on valence classification increased by 25 %.
Future works will deepen the hypothesis of a relationship between EDA and
these proprieties of olfactory stimulation and it will include other features and
physiological signals (like HRV) in the pattern recognition problem.
Mood Assessment in Bipolar Disorder In this study, EDA analysis was per-
formed in ten bipolar patients recruited in the frame of the European project
PSYCHE [8–10]. Each patient’s mood state was clinically evaluated as depressed,
euthymic or mixed. The patients were asked to passively view a set of IAPS images
and to describe TAT pictures. Novelties of this work are mainly related to data,
experimental protocol, and signal processing methodology. The innovative appli-
cation of EDA analysis allowed us to effectively test the experimental hypothesis
of having different sympathetic activations among different pathological mental
states. As a consequence, the proposed EDA feature set could have a prognostic
value on mental illness and can be evaluated when the SMNA is estimated using a
deconvolution model.
Both a deconvolution and a convex optimization analysis were applied to the
EDA signals in order to perform an effective separation of the EDA components
into tonic and phasic drivers. Several features were extracted in order to quantify and
characterize such components allowing for intra-subject and inter-subject statistical
analysis. The idea behind this study is that when a patient is depressed reacts
less intensively to high arousing stimuli than in experiencing mixed-state, while
sympathetic activity remarkably increases when the patient is in the euthymic state.
This is confirmed by Fig. 5.26a. Accordingly, Table 5.32 shows how the phasic
contribution increases or decreases during mood swings. The discordant trend of
118 6 Conclusions
Pz04 can be justified by the presence in the mixed-state of maniac symptoms,

even if the literature is quite poor on the relationship between maniac states
and EDA. Analysis of features extracted from the tonic CDA signal (stimulus-
unrelated component) during IAPS stimulation, revealed a significant difference in
the acquisitions of all the patients but Pz02 and Pz04. As a consequence, on the
basis of the limits of the results, no final conclusion can be drawn about a possible
link of this component and mood swings. Instead, cvxEDA model did not show
any significant differences between the euthymic and the other mood states in all
patients, suggesting that the tonic level was not a suitable marker to statistically
discriminate the mood state variation Statistical analyses were also intended as
inter-subject evaluations and performed using the Kruskal-Wallis non-parametric
test. Accordingly, post-hoc analysis engaged non-parametric Mann-Whitney tests
considering Bonferroni adjustment of the statistical significance. Grouping the
acquisition with the same label, the statistical analysis showed strong differences
among the three mood states under examination. Specifically, phasic features well
discriminated among depression, mixed state, and euthymia. An incremental trend
of the signal was observed over these three states. The depression condition is
confirmed to lead to a severe decrease of the electrodermal response activity and
consequently of the ANS activity. During the mixed-state phase, the patients exhibit
a higher level in the phasic activity, i.e. a stronger response to the stimuli, which
is, however, significantly lower than that seen in the euthymic state, in which the
subject feels like in normal conditions.
Differently, the tonic features regarding IAPS stimuli analyzed through CDA
showed a strong separation between the euthymic and mixed-state, which shows
a strong tonic hypoactivity. The tonic component, which is not directly connected to
the stimuli but is related to the state of the subject, showed an overlap between the
depressive and mixed state and the depressive and euthymia. No information was
gathered from the tonic components of the cvxEDA model.
Moreover, the relationship between mood state and EDA is confirmed by a
multivariate pattern recognition analysis. The three mood states under investigation
were discriminated with a high accuracy by a K-NN classifier. In the comparison
between CDA and cvxEDA models there are similar results. Both of them showed
an overall accuracy over 80 %. CDA has slightly better performances using both
features set, instead cvxEDA showed slightly better accuracy using the phasic
feature set.
Finally, results performed on healthy subjects strongly support the hypothesis
that EDA signal processing provides a viable decision support systems for mental
disorders. Healthy subjects, in fact, shown no statistical difference on each of the
EDA feature pattern between multiple affective elicitation along the time. Thus, it
is reasonable that the coherent changes found in the bipolar patients group can be
considered as real biomarkers of pathological mood states.
As a preliminary evaluation, it is necessary to mention some limitations in
this study. In fact, the whole study relies on the patient mood label given by the
physician during the training phase. Therefore, an error in such an evaluation could
be crucial for the further assessment biasing the decision support. In addition, more
6 Conclusions 119
data coming from a statistical representative and homogeneous population of a

bipolar patient is needed for the validation of the system in terms of generalization,
robustness and reliability.
In conclusion, results confirm the hypothesis of a link between changes in EDA
and mood states. Specifically, EDA strongly changed in the different mood states in
response to affective stimuli, showing a specific decrease in depressive phases. On
this basis, we conclude that EDA variations in phasic components can be suitable
markers for discriminating mood states in bipolar patients.
Study on the Source Oscillations of Skin Admittance We used a novel wearable
EDA acquisition system able to acquire the exosomatic EDA using both DC and AC
methods. In order to test the usability of the novel sensorized glove and to investigate
about possible differences between DC and AC stimulation (i.e. 10 Hz, 100 Hz,
1 kHz), we designed an experimental paradigm where 40 healthy subjects were
stimulated by means of a mechanical expiration task and visual affective stimuli
selected from the IAPS database. From this collection of pictures ranked in terms
of arousal and valence level, two groups of images were selected: a group of neutral
images and a group of negative aroused images.
The EDA signals were analyzed by means of the cvxEDA model [69]. The
cvxEDA algorithm is based on the three concepts of sparsity, Bayesian statistics and
convex optimization. It provides a decomposition of the EDA in its two components,
i.e. phasic and tonic, and estimates the sudomotor nerve activity that control the
eccrine sweat process, giving a window on the sympathetic nerve activity.
Results from the application of the cvxEDA algorithm showed no differences
in the identification of the phasic peak response after the deep respiration stimulus
among the DC and three AC methods. In fact, over 97 % of the peaks were identified
in the SMNA signal in the time response window of 5 s after the stimulus onset
(i.e., directly evoked by the stimulus [15, 16]). We could conclude that all of the
investigated methods for the exosomatic measurement of the EDA reliably measure
the phasic responses to eliciting stimuli.
Considering the four groups of data separately (i.e., DC, AC 10 Hz, AC 100 Hz,
AC 1 kHz), in the second part of the experiment we investigated possible differences
in inferring the arousal state. Specifically, we performed a classification procedure
of the arousal levels in the four groups of signal. Results showed that an alternating
current method at 100 Hz could improve the arousal recognition accuracy up to
71 % (while other acquisition modalities did not overcome an average accuracy
of 63.5 %). This results suggested that not only the skin conductance play an
important role in the electrodermal affective response, but also the susceptance (i.e.,
imaginary part of the skin impedance) may contain relevant information about the
SNS. Moreover, this relationship between AC frequency and recognition accuracy
is strongly nonlinear due to the nonlinear relationship between skin impedance, and
amplitude and frequency of the external electrical source [322]. Specifically, it is
well-known that the current density under a surface plate electrode could be not
uniform, and electrode surface presents fractal properties creating local areas of
different current densities. The onset of non-linearity may therefore be gradual,
120 6 Conclusions
and starts very early at very limited areas on the electrode surface (e.g., It has
been showed that very weak non-linearity is measurable at voltages than 100 mV).
Hence, it may be difficult to differentiate between the non-linearity of the electrode
processes and the tissue processes [323].
We are aware that works stated that the role of the susceptance is less important
with respect to the conductance at low frequency [40], but our results seem indicate
that a significant difference in EDA results are frequency dependent even more when
they are not mechanical but emotional evoked.
In other words, we assume that it could be feasible that emotional stimuli may
involve a capacitive component in the medium under investigation that has a bigger
contribution at 100 Hz.
Moreover, we should take into account that the Ohm’s law, given by J D E,
in such a medium could be not valid and it may be useful to treat as a complex
quantity in order to incorporate dielectric losses and frequency dependence [324],
therefore defining as: D 0 C j 00 (where j is the imaginary unit).
Future works will investigate the real and imaginary component of the admittance
in the analysis of the EDA dynamics by involving time varying method that could
highlight the nonlinear nature of the electrodermal response.
6.1 Future Challenges
The applicability of cvxEDA approach is not limited to EDA analysis but can be
extended to other domains requiring the deconvolution of pulse trains from the
output of systems that can be represented as ARMA models, for example in calcium
imaging [79] or hormone secretion analysis [82].
Moreover, to achieve a more complete description of the ANS and CNS activity,
both cvxEDA tonic and phasic outputs could be used to develop multivariate
models. For instance, the estimated SMNA component and EEG signals could be
related to each other via certain joint probabilities in order to assess joint CNS-
ANS dynamics especially related to sympathetic activity during, e.g., emotional
elicitation. Moreover, the tonic component, which can be considered as source of
reliable information on the sympathetic tone [85] (see also Sect. 2.5.2), could be
modeled using bivariate system of equations along with series of heartbeat dynamics
to provide information related to the sympatho-vagal balance.
Envisioned future challenging applications can be related to assistive devices
and rehabilitation, e.g. for patients with severe brain damages, which can be in
one of several states collectively known as Disorders Of Consciousness (DOC).
Indeed, the treatment of these patients is often driven by subjective experience, and
self-intuition of the clinicians. Nevertheless, there is not a standardized approach
to investigate if some perceptual channels (e.g., touch, which was proven to
communicate distinct emotions as discussed in the Introduction) are still active
in DOC. DOC assessment and rehabilitation could benefit from affective elicitation
6.1 Future Challenges 121
and recognition systems, like the ones described in this book. Finally, the cvxEDA
algorithm could be also profitably used in conjunction with wearable/portable
sensing systems (e.g., Empatica tools [325], or like the one in [42, 57]) to assess
mental and physical stress [326], which are key parameters to be monitored during
the course of a rehabilitation intervention.
Bibliography
[1] Benedek, M., & Kaernbach, C. (2010). A continuous measure of phasic electrodermal
activity. Journal of neuroscience methods, 190(1), 80–91.
[2] Matu, S., Cristea, I., Coteţ, C., Valenza, G., Gentili, C., Scilingo, E., et al. (2012). Are
socially anxious individuals less empathic? a psychophysiological investigation of facial
mimicry for emotional expressions. International Journal of Psychophysiology, 85(3), 375–
376.
[3] Cristea, I., Valenza, G., Gentili, C., Tatar, A., Scilingo, E., & David, D. (2012). Cognitive
reappraisal and acceptance distinctively impact heart rate variability in socially anxious
individuals. International Journal of Psychophysiology, 85(3), 339.
[4] Zajonc, R. (1984). On the primacy of affect. American Psychologist, 39(2), 117–123.
[5] Gross, J., & Muñoz, R. (1995). Emotion regulation and mental health. Clinical Psychology:
Science and Practice, 2(2), 151–164.
[6] Lazarus, R., & Averill, J. (1972). Emotion and cognition: With special reference to anxiety.
Anxiety: Current Trends in Theory and Research, 2, 242–284.
[7] Damasio, A. (2000). Descartes’ error: Emotion, reason, and the human brain. New York:
Quill.
[8] Valenza, G., Gentili, C., Lanata, A., & Scilingo, E. (2013). Mood recognition in bipolar
patients through the psyche platform: preliminary evaluations and perspectives. Artificial
Intelligence In Medicine, 57(1), 49–58.
[9] Greco, A., Lanata, A., Valenza, G., Rota, G., Vanello, N., & Scilingo, E. (2012). On
the deconvolution analysis of electrodermal activity in bipolar patients. In 2012 Annual
International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)
(pp. 6691–6694). IEEE.
[10] Vanello, N., Guidi, A., Gentili, C., Werner, S., Bertschy, G., Valenza, G., et al. (2012).
Speech analysis for mood state characterization in bipolar patients. In 2012 Annual
International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)
(pp. 2104–2107). IEEE.
[11] Greco, A., Valenza, G., Lanata, A., Rota, G., & Scilingo, E. P. (2014). Electrodermal activity
in bipolar patients during affective elicitation. IEEE Journal of Biomedical and Health
Informatics, 18(6), 1865–1873.
[12] Fowles, D., Christie, M., Edelberg, R., Grings, W., Lykken, D., & Venables, P. (1981).
Publication recommendations for electrodermal measurements, Psychophysiology, 18(3),
232–239.
[13] Edelberg, R. (1972). Electrical activity of the skin: Its measurement and uses in psychophys-
iology. Handbook of psychophysiology (Vol. 12, p. 1011). New York: Holt.

with Applications for Mental Health, DOI 10.1007/978-3-319-46705-4
124 Bibliography
[14] Edelberg, R. (1993). Electrodermal mechanisms: A critique of the two-effector hypothesis

and a proposed replacement. In Progress in electrodermal research (pp. 7–29). Berlin:
Springer.
[15] Boucsein, W. (2012). Electrodermal activity (2nd ed). New York: Springer Science &
Business Media.
[16] Benedek, M., & Kaernbach, C. (2010). Decomposition of skin conductance data by means
of nonnegative deconvolution. Psychophysiology, 47(4), 647–658.
[17] Dawson, M. E., Schell, A. M., & Filion, D. L. (2007). The electrodermal system. In
J. T. Cacioppo, L. G. Tassinary, & G. G. Berntson (Eds.), Handbook of psychophysiology
(Chap. 7). Cambridge: Cambridge University Press.
[18] Schmidt, S., & Walach, H. (2000). Electrodermal activity (EDA)–state-of-the-art measure-
ments and techniques for parapsychological purposes. The Journal of Parapsychology, 64(2),
139.
[19] Roth, W. T., Dawson, M. E., & Filion, D. L. (2012). Publication recommendations for
electrodermal measurements. Psychophysiology, 49, 1017–1034.
[20] Greco, A., Lanata, A., Valenza, G., Rota, G., Vanello, N., & Scilingo, E. P. (2012). On the
deconvolution analysis of electrodermal activity in bipolar patients. Proceedings of the 34th
IEEE EMBS Conference (Vol. 2012, pp. 6691–6694).
[21] Greco, A., Valenza, G., Lanata, A., Rota, G., & Scilingo, E. P. (2014). Electrodermal activity
in bipolar patients during affective elicitation. IEEE Journal of Biomedical and Health
Informatics, 18(6), 1865–1873.
[22] Venables, P. H., & Christie, M. J. (1980). Electrodermal activity. Techniques in Psychophys-
iology, 54(3), 3–67.
[23] Boucsein, W. (1992). Electrodermal activity (2nd ed.). New York: Springer Science &
Business Media.
[24] Millington, P. F., & Wilkinson, R. (1983). Skin. In R. J. Harrison, & R. M. McMinn (Series
Eds.) Biological structure and function (Vol. 9, pp. 83–98). Cambridge (Great Britain):
Cambridge University Press.
[25] Tregear, R. T. (1966). Physical functions of skin (Vol. 5). London: Academic Press.
[26] Walters, K. A., & Roberts, M. S. (2002). The structure and function of skin. Drugs and the
Pharmaceutical Sciences, 119, 1–40.
[27] Quay, W. (1977). Structure and function of skin glands. In Chemical signals in vertebrates
(pp. 1–16). Berlin: Springer.
[28] Kuno, Y.(1956). Human perspiration (no. 285). Illinois: Thomas.
[29] Pinkus, H. (1952). Examination of the epidermis by the strip method. Journal of Investiga-
tive Dermatology, 19(6), 431–447.
[30] Herrmann, F., & Ippen, H. (1973). Biochemie der Haut:: 71 Tab. Thieme. Stuttgart.
[31] Montagna, W. (2012). The structure and function of skin 3E. New York: Academic.
[32] Sato, K. (1983). The physiology and pharmacology of the eccrine sweat gland. Biochemistry
and Physiology of the Skin, 1, 569–595.
[33] Martin, I. (1980). Techniques in psychophysiology (Vol. 6). New York: Wiley.
[34] Fowles, D. C. (1986). The eccrine system and electrodermal activity. Psychophysiology:
Systems, Processes, and Applications, 1, 51–96.
[35] Bagshaw, M. H., Kimble, D. P., & Pribram, K. H. (1965). The gsr of monkeys during ori-
enting and habituation and after ablation of the amygdala, hippocampus and inferotemporal
cortex. Neuropsychologia, 3(2), 111–119.
[36] Montagu, J., & Coles, E. (1966). Mechanism and measurement of the galvanic skin response.
Psychological Bulletin, 65(5), 261.
[37] Edelberg, R. (1983). The effects of initial levels of sweat duct filling and skin hydration on
electrodermal response amplitude. Psychophysiology, 20(5), 550–557.
[38] Boucsein, W., Schaefer, F., & Neijenhuisen, H. (1989). Continuous recordings of impedance
and phase angle during electrodermal reactions and the locus of impedance change.
Psychophysiology, 26(3), 369–376.
Bibliography 125
[39] Martínez-Rodrigo, A., Zangróniz, R., Pastor, J. M., & Fernández-Caballero, A. (2015).
Arousal level classification in the ageing adult by measuring electrodermal skin conductivity.
In Ambient intelligence for health (pp. 213–223). New York: Springer.
[40] Martinsen, Ø., Grimnes, S., & Sveen, O. (1997). Dielectric properties of some keratinised
tissues. part 1: Stratum corneum and nail in situ. Medical and Biological Engineering and
Computing, 35(3), 172–176.
[41] Hanson, M. A., Powell, Jr. H. C., Barth, A. T., Ringgenberg, K., Calhoun, B. H., Aylor, J. H.,
et al. (2009). Body area sensor networks: Challenges and opportunities. Computer, 42(1), 58.
[42] Carbonaro, N., Greco, A., Anania, G., Dalle Mura, G., Tognetti, A., Scilingo, E., et al.
(2012). Unobtrusive physiological and gesture wearable acquisition system: A preliminary
study on behavioral and emotional correlations. Global Health, 1, 88–92.
[43] Lee, Y., Lee, B., & Lee, M. (2010). Wearable sensor glove based on conducting fabric using
electrodermal activity and pulse-wave sensors for e-health application. Telemedicine and
e-Health, 16(2), 209–217.
[44] Patel, S., Park, H., Bonato, P., Chan, L., & Rodgers, M. (2012). A review of wearable
sensors and systems with application in rehabilitation. Journal of Neuroengineering and
Rehabilitation, 9(1), 1.
[45] Garbarino, M., Lai, M., Bender, D., Picard, R. W., & Tognetti, S. (2014). Empatica e3—
a wearable wireless multi-sensor device for real-time computerized biofeedback and data
acquisition. In 2014 EAI 4th International Conference on Wireless Mobile Communication
and Healthcare (Mobihealth) (pp. 39–42). IEEE.
[46] Martinsen, O., & Grimnes, S. (1998). On using single frequency electrical measurements for
skin hydration assessment. Innovation et Technologie en Biologie et Médecine, 19, 395–400.
[47] Martinsen, Ø. G., & Grimnes, S. (2001). Facts and myths about electrical measurement of
stratum corneum hydration state. Dermatology, 202(2), 87–89.
[48] Martinsen, Ø. G., Grimnes, S., Nilsen, J. K., Tronstad, C., Jang, W., Kim, H. et al.
(2008). Gravimetric method for in vitro calibration of skin hydration measurements. IEEE
Transactions on Biomedical Engineering, 55(2), 728–732.
[49] Picard, R. (2010). Emotion research by the people, for the people. Emotion Review, 2(3),
250.
[50] Marieke van Doorena, J.J.G. (Gert-Jan) de Vriesa, & Janssena, J. H. (2012). Emotional
sweating across the body: Comparing 16 different skin conductance measurement locations.
Physiology & Behavior, 106(2), 298–304.
[51] Poh, M.-Z., Swenson, N. C., & Picard, R. W. (2010). A wearable sensor for unobtrusive,
long-term assessment of electrodermal activity. IEEE Transactions on Biomedical Engineer-
ing, 57(5), 1243–1252.
[52] Strauss, M., Reynolds, C., Hughes, S., Park, K., McDarby, G., & Picard, R. W. (2005).
The handwave bluetooth skin conductance sensor. In International Conference on Affective
Computing and Intelligent Interaction (pp. 699–706). Springer.
[53] Thought Technology Ltd: ProComp Infinity Encoder and amplifiers. http://
thoughttechnology.com/index.php/hardware/flexcomp-system-with-biograph-infiniti-
software-t7555m.html (2016).
[54] Biopac Systems, Inc. http://www.biopac.com/ (2016).
[55] BioSemi Instrumentation. http://www.biosemi.com/ (2016).
[56] Scheirer, J., & Picard, R. (2001). The Galvactivator: A glove that senses and communicates
skin conductivity. In Proceedings of the 9th International Conference on Human-Computer
Interaction.
[57] Lanata, A., Valenza, G., & Scilingo, E. (2012). A novel EDA glove based on textile-
integrated electrodes for affective computing. Medical and Biological Engineering and
Computing, 50, 1163–1172.
[58] Brainquiry, B. V. (2005) PET-GSR Wireless. http://www.brainquiry.com/.
[59] Analog Device: AD9833 Low Power, Programmable Waveform Generator. http://www.
analog.com/en/products/rf-microwave/direct-digital-synthesis-modulators/ad9833.html
(2016).
126 Bibliography
[60] Texas Instrument: MSP430 ultra-low-power Microcontrollers, http://www.ti.com/lsds/ti/

microcontrollers_16-bit_32-bit/msp/overview.page (2016).
[61] Breska, A., Maoz, K., & Ben-Shakhar, G. (2011). Interstimulus intervals for skin conduc-
tance response measurement. Psychophysiology, 48(4), 437–440.
[62] Barry, R. J., Feldmann, S., Gordon, E., Cocker, K. I., & Rennie, C. (1993). Elicitation
and habituation of the electrodermal orienting response in a short interstimulus interval
paradigm. International Journal of Psychophysiology, 15(3), 247–253.
[63] Lim, C. L., Rennie, C., Barry, R. J., Bahramali, H., Lazzaro, I., Manor, B., et al. (1997).
Decomposing skin conductance into tonic and phasic components. International Journal of
[64] Bach, D. R., & Friston, K. J. (2013). Model-based analysis of skin conductance responses:
Towards causal models in psychophysiology. Psychophysiology, 50(1), 15–22.
[65] Alexander, D., Trengove, C., Johnston, P., Cooper, T., August, J., & Gordon, E. (2005).
Separating individual skin conductance responses in a short interstimulus-interval paradigm.
Journal of Neuroscience Methods, 146(1), 116–123.
[66] Bach, D. R. (2014). A head-to-head comparison of SCRalyze and Ledalab, two model-based
methods for skin conductance analysis. Biological Psychology, 103, 63–68.
[67] Chaspari, T., Tsiartas, A., Stein, L., Cermak, S., & Narayanan, S. (2015). Sparse represen-
tation of electrodermal activity with knowledge-driven dictionaries. IEEE Transactions on
Biomedical Engineering, 62(3), 960–971.
[68] Boyd, S. P., & Vandenberghe, L. (2004) Convex optimization. Cambridge: Cambridge
University Press.
[69] Greco, A., Valenza, G., Lanata, A., Scilingo, E., & Citi, L. (2016). cvxEDA: A convex opti-
mization approach to electrodermal activity processing. IEEE Transactions on Biomedical
Engineering, 63(4), 797–804.
[70] Karenbach, C. (2005). Ledalab-a software package for the analysis of phasic electrodermal
activity. Technical Report, Allgemeine Psychologie, Institut für Psychologie.
[71] Gustafsson, F. (1996). Determining the initial states in forward-backward filtering. IEEE
Transactions on Signal Processing, 44(4), 988–992.
[72] Mitra, S. (2001). Digital signal processing. a computer-based approach. New York:
McGraw-Hill
[73] Schneider, R. (1987). A mathematical model of human skin conductance. Psychophysiology,
24(5), 610.
[74] Garrett, E. (1994). The Bateman function revisited: A critical reevaluation of the quantitative
expressions to characterize concentrations in the one compartment body model as a function
of time with first-order invasion and first-order elimination. Journal of Pharmacokinetics
and Pharmacodynamics, 22(2), 103–128.
[75] Macefield, V., & Wallin, B. (1996). The discharge behaviour of single sympathetic neurones
supplying human sweat glands. Journal of the Autonomic Nervous System, 61(3), 277–286.
[76] Nishiyama, T., Sugenoya, J., Matsumoto, T., Iwase, S., & Mano, T. (2001). Irregular
activation of individual sweat glands in human sole observed by a videomicroscopy.
Autonomic Neuroscience, 88(1), 117–126.
[77] Bach, D. R., Flandin, G., Friston, K. J., & Dolan, R. J. (2010). Modelling event-related skin
conductance responses. International Journal of Psychophysiology, 75(3), 349–356.
[78] Bach, D. R., Friston, K. J., & Dolan, R. J. (2013). An improved algorithm for model-based
analysis of evoked skin conductance responses. Biological Psychology, 94(3), 490–497.
[79] Vogelstein, J. T., Packer, A. M., Machado, T. A., Sippy, T., Babadi, B., Yuste, R. et al. (2010).
Fast nonnegative deconvolution for spike train inference from population calcium imaging.
Journal of Neurophysiology, 104(6), 3691–3704.
[80] O’Brien, M. S., Sinclair, A. N., & Kramer, S. M. (1994). Recovery of a sparse spike
time series by L1 norm deconvolution. IEEE Transactions on Signal Processing, 42(12),
3353–3365.
[81] Tibshirani, R. (1996). Regression shrinkage and selection via the LASSO. Journal of the
Royal Statistical Society. Series B (Methodological), 58, 267–288.
Bibliography 127
[82] de Rooi, J., & Eilers, P. (2011). Deconvolution of pulse trains with the L0 penalty. Analytica
Chimica Acta, 705(1), 218–226.
[83] Levinson, D., & Edelberg, R. (1985). Scoring criteria for response latency and habituation
in electrodermal research: a critique. Psychophysiology, 22(4), 417–426.
[84] Ishchenko, A., & Shev’ev, P. (1989). Automated complex for multiparameter analysis of the
galvanic skin response signal. Biomedical Engineering, 23(3), 113–117.
[85] Posada-Quintero, H. F., Florian, J. P., Orjuela-Cañón, A. D., Aljama-Corrales, T.,
Charleston-Villalobos, S., & Chon, K. H. (2016). Power spectral density analysis of elec-
trodermal activity for sympathetic function assessment. Annals of Biomedical Engineering,
44, 1–12.
[86] Grassi, G., & Esler, M. (1999). How to assess sympathetic activity in humans. Journal of
Hypertension, 17(6), 719–734.
[87] Bauer, A., Malik, M., Schmidt, G., Barthel, P., Bonnemeier, H., Cygankiewicz, I., et al.
(2008). Heart rate turbulence: Standards of measurement, physiological interpretation, and
clinical use: International society for Holter and noninvasive electrophysiology consensus.
Journal of the American College of Cardiology, 52(17), 1353–1365.
[88] Crider, A., & Lunn, R. (1971). Electrodermal lability as a personality dimension. Journal of
Experimental Research in Personality, 5(2), 145–150.
[89] Kira, Y., Ogura, T., Aramaki, S., Kubo, T., Hayasida, T., & Hirasawa, Y. (2001). Sympathetic
skin response evoked by respiratory stimulation as a measure of sympathetic function.
Clinical Neurophysiology, 112(5), 861–865.
[90] Lang, P., Bradley, M., & Cuthbert, B. (2005). International affective picture system iaps):
Digitized photographs, instruction manual and affective ratings. Technical Report A-6.
University of Florida.
[91] Cook, E. W., Hawk, L. W., Davis, T. L., & Stevenson, V. E. (1991). Affective individual
differences and startle reflex modulation. Journal of Abnormal Psychology, 100(1), 5.
[92] Lang, P., Greenwald, M., Bradley, M., & Hamm, A. (1993). Looking at pictures: Affective,
facial, visceral, & behavioral reactions. Psychophysiology, 30(3), 261–273.
[93] Hollander, M., Wolfe, D. A., & Chicken, E. (2014). Nonparametric statistical methods. New
York: Wiley.
[94] Frijda, N. (1986). The emotions. Cambridge: Cambridge University Press.
[95] Ekman, P. E., & Davidson, R. J. (1994). The nature of emotion: Fundamental questions.
Oxford: Oxford University Press.
[96] Russell, J., & Carroll, J.M. (1999). On the bipolarity of positive and negative affect.
Psychological Bulletin, 125(1), 3–30.
[97] Watson, D., Wiese, D., Vaidya, J., & Tellegen, A. (1999). The two general activation systems
of affect: Structural findings, evolutionary considerations, and psychobiological evidence.
Journal of Personality and Social Psychology, 76(5), 820–838.
[98] Watson, D., & Clark, L. (1992). On traits and temperament: General and specific factors
of emotional experience and their relation to the five-factor model. Journal of Personality,
60(2), 441–476.
[99] LeDoux, J. (1998). The emotional brain: The mysterious underpinnings of emotional life.
New York: Simon and Schuster.
[100] Pegna, A. J., Khateb, A., Lazeyras, F., & Seghier, M. L. (2005). Discriminating emotional
faces without primary visual cortices involves the right amygdala. Nature Neuroscience,
8(1), 24–25.
[101] Ekman, P. (1993). Facial expression and emotion. American psychologist, 48(4), 384.
[102] Darwin, C. (1872). The expression of the emotions in man and animals; with an introduction,
afterword, and commentaries by Paul Ekman. New York: Oxford University.
[103] Ekman, P. (1974). 1. universal facial expressions of emotion. In R. A. LeVine (Ed.), Culture
and personality: Contemporary readings (pp. 8–15). Chicago: Aldine.
[104] Ekman, P. (1999). Basic emotions. In T. Dalgleish, & T. Power (Eds.), The Handbook of
cognition and emotion, (pp. 45–60). Sussex: Wiley.
128 Bibliography
[105] Tompkins, S. (1962). Affect Imagery Consciousness: Volume I: The Positive Affects.
New York: Springer Publishing Company.
[106] Izard, C. (1971). The face of emotion (Vol. 23). New York: Appleton-Century-Crofts.
[107] Plutchik, R. (1984). Emotions: A general psychoevolutionary theory. Approaches to Emo-
tion, 197–219.
[108] Ekman, P. (1973). Cross-cultural studies of facial expression. In P. Ekman (Ed.), Darwin and
facial expression: A century of research in review, (pp. 169–222). New York: Academic.
[109] Watson. J. (1997) Behaviorism. New Brunswick: Transaction Publication.
[110] Ortony, A., & Turner, T. (1990). What is basic about basic emotions. Psychological Review,
97(3), 315–331.
[111] Wundt, W. (1905). Grundriss der psychologie [Fundamentals of psychology]. (7th rev. ed.).
Liepzig: Engelman.
[112] Schlosberg, H. (1954). Three dimensions of emotion. Psychological Review, 61(2), 81–88.
[113] Osgood, C. (1975). The measurement of meaning. Champaign: University of Illinois Pr.
[114] Davitz, J. (1969). The language of emotion. New York: Academic.
[115] Lang, P., Bradley, M., & Cuthbert, B. (1998). Emotion, motivation, and anxiety: Brain
mechanisms and psychophysiology. Biological Psychiatry, 44(12), 1248–1263.
[116] Panskepp, J. (1998). Affective neuroscience: The foundations of human and animal emo-
tions. Oxford: Oxford University Press.
[117] Breazeal, C. (2003). Emotion and sociable humanoid robots. International Journal of
Human-Computer Studies, 59(1–2), 119–155.
[118] Russell, J. (1980). A circumplex model of affect. Journal of Personality and Social
Psychology, 39(6), 1161–1178.
[119] Russell, J., & Mehrabian, A. (1977). Evidence for a three-factor theory of emotions* 1.
Journal of Research in Personality, 11(3), 273–294.
[120] Arnold, M. B. (1950). An excitatory theory of emotion. In M. L. Reymert (Ed.), Feelings
and emotions (pp. 11–33). New York: McGraw-Hill.
[121] Ortony, A., Clore, G., & Collins, A. (1990). The cognitive structure of emotions. Cambridge:
Cambridge University Press.
[122] Scherer, K., & Ekman, P. (1984). Approaches to emotions. Londres et New Jersey: Lawrence
Erlbaum Associates, Publication.
[123] Lisetti, C., & Gmytrasiewicz, P. (2002). Can a rational agent afford to be affectless? A formal
approach. Applied Artificial Intelligence, 16, 1–33.
[124] Scherer, K., Schorr, A., & Johnstone, T. (2001). Appraisal processes in emotion: Theory,
methods, research. Oxford: Oxford University Press.
[125] Egges, A., Kshirsagar, S., & Magnenat-Thalmann, N. (2003). A model for personality and
emotion simulation. In Knowledge-based intelligent information and engineering systems
(pp. 453–461). Berlin: Springer.
[126] Posner, J., Russell, J., & Peterson, B. (2005). The circumplex model of affect: An
integrative approach to affective neuroscience, cognitive development, and psychopathology.
Development and Psychopathology, 17(03), 715–734.
[127] Nasoz, F., Alvarez, K., Lisetti, C., & Finkelstein, N. (2003). Emotion recognition from phys-
iological signals using wireless sensors for presence technologies. Cognition, Technology
and Work, 6, 4–14.
[128] Janig, W. (1989). Autonomic nervous system. In R. F. Schmidt, & G. Thews (Eds.), Human
physiology, (2nd ed.). Berlin: Springer.
[129] Lewis, M., Haviland-Jones, J. M., & Barrett, L. F. (2010). Handbook of emotions. New York:
Guilford Press.
[130] Picard, R. (2000). Affective computing. Cambridge: MIT.
[131] Picard, R., Vyzas, E., & Healey, J. (2001). Toward machine emotional intelligence: Analysis
of affective physiological state. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 23(10), 1175–1191.
Bibliography 129
[132] Kim, J., & André, E. (2008). Emotion recognition based on physiological changes in music
listening. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(12), 2067–
2083.
[133] Katsis, C. D., Katertsidis, N. S., & Fotiadis, D. I. (2010). An integrated system based on
physiological signals for the assessment of affective states in patients with anxiety disorders.
Biomedical Signal Processing and Control, 6(3), 261–268.
[134] Valenza, G., Greco, A., Citi, L., Bianchi, M., Barbieri, R., & Scilingo, E. (2016). Inhomoge-
neous point-processes to instantaneously assess affective haptic perception through heartbeat
dynamics information. Scientific Reports, 6(28567), 1–14.
[135] Valenza, G., Greco, A., Gentili, C., Lanata, A., Sebastiani, L., Menicucci, D., et al. (2016).
Combining electroencephalographic activity and instantaneous heart rate for assessing
brain–heart dynamics during visual emotional elicitation in healthy subjects. Philosophical
Transactions of the Royal Society A , 374(2067), 20150176.
[136] Valenza, G., Nardelli, M., Gentili, C., Bertschy, G., Kosel, M., Scilingo, E. P., et al.
(2016). Predicting mood changes in bipolar disorder through heartbeat nonlinear dynamics.
Biomedical and Health Informatics, 20(2), 1034–1043.
[137] Valenza, G., Citi, L., Gentili, C., Lanata, A., Scilingo, E. P., & Barbieri, R. (2014). Point-
process nonlinear autonomic assessment of depressive states in bipolar patients. Methods of
Information in Medicine, 53(4), 296–302.
[138] Valenza, G., Lanatà, A., Scilingo, E. P., & De Rossi, D. (2010). Towards a smart glove:
Arousal recognition based on textile electrodermal response. In 2010 Annual International
Conference of the IEEE Engineering in Medicine and Biology (pp. 3598–3601). IEEE.
[139] Mazzei, D., Greco, A., Lazzeri, N., Zaraki, A., Lanatà, A., Igliozzi, R., et al. (2012). Robotic
social therapy on children with autism: preliminary evaluation through multi-parametric
analysis. In 2012 International Conference on and 2012 International Confernece on Social
Computing (SocialCom) Privacy, Security, Risk and Trust (PASSAT) (pp. 955–960). IEEE.
[140] Betella, A., Zucca, R., Cetnarski, R., Greco, A., Lanatà, A., Mazzei, D., et al. (2015).
Inference of human affective states from psychophysiological measurements extracted under
ecologically valid conditions. Using Neurophysiological Signals that Reflect Cognitive or
Affective State, 8, 66.
[141] Lanatá, A., Valenza, G., Mancuso, C., & Scilingo, E. (2011). Robust multiple cardiac
arrhythmia detection through bispectrum analysis. Expert Systems with Applications, 38(6),
6798–6804.
[142] Valenza, G., & Scilingo, E. P. (2014) Autonomic nervous system dynamics for mood and
emotional-state recognition. Springer.
[143] James, W. (1884). II.—what is an emotion? Mind, os-IX (34), 188–205. doi:10.1093/
mind/os-IX.34.188.
[144] Ellsworth, P. C. (1994). William James and emotion: is a century of fame worth a century of
misunderstanding? Psychological Review, 101(2), 222.
[145] Lang, P. J. (1994). The varieties of emotional experience: A meditation on james-lange
theory. Psychological Review, 101(2), 211.
[146] Scherer, K. R., & Wallbott, H. G. (1994). Evidence for universality and cultural variation
of differential emotion response patterning. Journal of Personality and Social Psychology,
66(2), 310.
[147] Stemmler, G., Heldmann, M., Pauls, C. A., & Scherer, T. (2001). Constraints for emotion
specificity in fear and anger: The context counts. Psychophysiology, 38(2), 275–291.
[148] Ekman, P., Levenson, R., & Friesen, W. (1983). Autonomic nervous system activity
distinguishes among emotions. Science, 221(4616), 1208–1210.
[149] Christie, I., & Friedman, B. (2004). Autonomic specificity of discrete emotion and
dimensions of affective space: A multivariate approach. International Journal of Psy-
chophysiology, 51(2), 143–153.
[150] Cacioppo, J. T., Berntson, G. G., Larsen, J. T., Poehlmann, K. M., Ito, T. A., et al. (2000).
The psychophysiology of emotion. Handbook of Emotions, 2, 173–191.
[151] Arnold, M. B. (1960). Emotion and personality. New York: Columbia University.
130 Bibliography
[152] Hillman, J. (1960). Emotion: a comprehensive phenomenology of theories and their meaning
for therapy. Evanston: Northwestern University Press.
[153] Campos, J. J., & Johnson, H. J. (1967). Affect, verbalization, and directional fractionation
of autonomic responses. Psychophysiology, 3(3), 285–290.
[154] Bradley, M. M., & Lang, P. J. (2000). Measuring emotion: Behavior, feeling, and physiology.
Cognitive Neuroscience of Emotion, 25, 49–59.
[155] Stemmler, G. (2004). Physiological processes during emotion. In P. Philippot & R. S.
Feldman (Eds.), The regulation of emotion (pp. 33–70). Mahwah: Erlbaum.
[156] Calvo, R., & D’Mello, S. (2010). Affect detection: An interdisciplinary review of models,
methods, and their applications. IEEE Transactions on Affective Computing, 1(1), 18–37.
[157] Greco, A., Valenza, G., Nardelli, M., Lanata, A., Bianchi, M., & Scilingo, E. P. (2015). Elec-
trodermal activity analysis during affective haptic elicitation. In 2015 Annual International
Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE.
[158] Swangnetr, M., & Kaber, D. B. (2012). Emotional state classification in patient–robot
interaction using wavelet analysis and statistics-based feature selection. IEEE Transactions
on Systems, Man and Cybernetics, 43(1), 63–75.
[159] Katsis, C., Katertsidis, N., Ganiatsas, G., & Fotiadis, D. (2008). Toward emotion recognition
in car-racing drivers: A biosignal processing approach. IEEE Transactions on Systems, Man
and Cybernetics, Part A: Systems and Humans, 38(3), 502–512.
[160] Lanatà, A., Valenza, G., Greco, A., Gentili, C., Bartolozzi, R., Bucchi, F., et al. (2015).
How the autonomic nervous system and driving style change with incremental stressing
conditions during simulated driving. IEEE Transactions on Intelligent Transportation
Systems, 16(3), 1505–1517.
[161] Chakraborty, A., Konar, A., Chakraborty, U., & Chatterjee, A. (2009). Emotion recognition
from facial expressions and its control using fuzzy logic. IEEE Transactions on Systems,
Man and Cybernetics, Part A: Systems and Humans, 39(4), 726–743.
[162] Chanel, G., Rebetez, C., Bétrancourt, M., & Pun, T. (2011). Emotion assessment from
physiological signals for adaptation of game difficulty. IEEE Transactions on Systems, Man
and Cybernetics, Part A: Systems and Humans, 41(6), 1052–1063.
[163] Lisetti, C., & Nasoz, F. (2004). Using noninvasive wearable computers to recognize human
emotions from physiological signals. EURASIP Journal on Applied Signal Processing, 2004,
1672–1687.
[164] Haag, A., Goronzy, S., Schaich, P., & Williams, J. (2004). Emotion recognition using bio-
sensors: First steps towards an automatic system. In Tutorial and research workshop on
affective dialogue systems. Lecture Notes in Computer Science (Vol. 3068, pp. 36–48).
Berlin/Heidelberg: Springer.
[165] Kim, K., Bang, S., & Kim, S. (2004). Emotion recognition system using short-term
monitoring of physiological signals. Medical and Biological Engineering and Computing,
42(3), 419–427.
[166] Yoo, S., Lee, C., Park, Y., Kim, N., Lee, B., & Jeong, K. (2005). Neural network
based emotion estimation using heart rate variability and skin resistance. In International
conference on natural computation. Lecture Notes in Computer Science (Vol. 3610,
pp. 818–824). Berlin/Heidelberg: Springer.
[167] Choi, A., & Woo, W. (2005). Physiological sensing and feature extraction for emotion
recognition by exploiting acupuncture spots. In International conference on affective
computing and intelligent interaction. Lecture Notes in Computer Science (Vol. 3784,
[168] Healey, J., & Picard, R. (2005). Detecting stress during real-world driving tasks using
physiological sensors. IEEE Transactions on Intelligent Transportation Systems, 6(2),
156–166.
[169] Li, L., & Chen, J. (2006). Emotion recognition using physiological signals. In Advances
in artificial reality and tele-existence. Lecture Notes in Computer Science (Vol. 4282,
Bibliography 131
[170] Rani, P., Liu, C., Sarkar, N., & Vanman, E. (2006). An empirical study of machine
learning techniques for affect recognition in human–robot interaction. Pattern Analysis &
Applications, 9(1), 58–69.
[171] Rainville, P., Bechara, A., Naqvi, N., & Damasio, A. (2006). Basic emotions are associated
with distinct patterns of cardiorespiratory activity. International Journal of Psychophysiol-
ogy, 61(1), 5–18.
[172] Zhai, J., & Barreto, A. (2006). Stress detection in computer users based on digital signal
processing of noninvasive physiological variables. In 28th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (pp. 1355–1358).
[173] Leon, E., Clarke, G., Callaghan, V., & Sepulveda, F. (2007). A user-independent real-time
emotion recognition system for software agents in domestic environments. Engineering
Applications of Artificial Intelligence, 20(3), 337–345.
[174] Liu, C., Conn, K., Sarkar, N., & Stone, W. (2008). Physiology-based affect recognition for
computer-assisted intervention of children with Autism Spectrum Disorder. International
Journal of Human-Computer Studies, 66(9), 662–677.
[175] Yannakakis, G., & Hallam, J. (2008). Entertainment modeling through physiology in
physical play. International Journal of Human-Computer Studies, 66(10), 741–755.
[176] Gouizi, K., Bereksi Reguig, F., & Maaoui, C. (2011). Emotion recognition from physiolog-
ical signals. Journal of Medical Engineering & Technology, 35(6–7), 300–307.
[177] Valenza, G., Lanatà, A., & Scilingo, E. (2011). The role of nonlinear dynamics in affective
valence and arousal recognition. IEEE Transactions on Affective Computing. doi:10.1109/T-
AFFC.2011.30.
[178] Jang, E.-H., Park, B.-J., Kim, S.-H., & Sohn, J.-H. (2012). Emotion classification based on
physiological signals induced by negative emotions: Discriminantion of negative emotions
by machine learning algorithm. In 2012 9th IEEE International Conference on Networking,
Sensing and Control (ICNSC) (pp. 283–288). IEEE.
[179] Ren, P., Barreto, A., Gao, Y., & Adjouadi, M. (2011). Comparison of the use of pupil
diameter and galvanic skin response signals for affective assessment of computer users.
Biomedical Sciences Instrumentation, 48, 345–350.
[180] Torres, C. A., Orozco, Á. A., & Alvarez, M. A. (2013). Feature selection for multimodal
emotion recognition in the arousal-valence space. In 2013 35th Annual International Con-
ference of the IEEE Engineering in Medicine and Biology Society (EMBC) (pp. 4330–4333).
IEEE.
[181] Zhaofang Y., & Guangyuan, L. (2013). Emotion recognition based on nonlinear features
of skin conductance response. Journal of Information & Computational Science, 10(12),
3877–3887.
[182] Maaoui, C., Abdat, F., & Pruski, A. (2014). Physio-visual data fusion for emotion recogni-
tion. IRBM, 35(3), 109–118.
[183] Kukolja, D., Popović, S., Horvat, M., Kovač, B., & Ćosić, K. (2014). Comparative
analysis of emotion estimation methods based on physiological measurements for real-time
applications. International Journal of Human-Computer Studies, 72(10), 717–727.
[184] Hariharan, A., & Adam, M. T. P. (2015). Blended emotion detection for decision support.
IEEE Transactions on Human-Machine Systems, 45(4), 510–517.
[185] Khezri, M., Firoozabadi, M., & Sharafat, A. R. (2015). Reliable emotion recognition system
based on dynamic adaptive fusion of forehead biopotentials and physiological signals.
Computer Methods and Programs in Biomedicine, 122(2), 149–164.
[186] Khan, A. M., & Lawo, M. (2016). Recognizing emotion from blood volume pulse and
skin conductance sensor using machine learning algorithms (pp. 1291–1297). Cham:
Springer International Publishing. Available: http://dx.doi.org/10.1007/978-3-319-32703-
7_247 [Online].
[187] Rukavina, S., Gruss, S., Hoffmann, H., Tan, J.-W., Walter, S., & Traue, H. C. (2016).
Affective computing and the impact of gender and age. PloS One, 11(3), e0150584.
132 Bibliography
[188] Bialoskorski, L. S., Westerink, J. H., & Broek, E. L. (2009). Mood swings: An affective
interactive art system. In International conference on intelligent technologies for interactive
entertainment. Lecture Notes of the Institute for Computer Sciences, Social Informatics and
Telecommunications Engineering (Vol. 9, pp. 181–186). Berlin/Heidelberg: Springer.
[189] Lisetti, C. L., & Nasoz, F. (2004). Using noninvasive wearable computers to recognize
human emotions from physiological signals. EURASIP Journal on Advances in Signal
Processing, 2004(11), 1–16.
[190] Nardelli, M., Valenza, G., Greco, A., Lanata, A., & Scilingo, E. (2015). Recognizing
emotions induced by affective sounds through heart rate variability. IEEE Transactions of
Affective Computing, 6(4), 385–394.
[191] Janssen, J., Van den Broek, E., & Westerink, J. (2009). Personalized affective music player.
In 3rd International Conference on Affective Computing and Intelligent Interaction and
Workshops, 2009 (ACII 2009) (pp. 1–6.). IEEE.
[192] Lin, Y., Wang, C., Jung, T., Wu, T., Jeng, S., Duann, J., & Chen, J. (2010). Eeg-based
emotion recognition in music listening. IEEE Transactions on Biomedical Engineering,
57(7), 1798–1806.
[193] Wagner, J., Kim, J., & André, E. (2005). From physiological signals to emotions: Imple-
menting and comparing selected methods for feature extraction and classification. In 2005
IEEE International Conference on Multimedia and Expo (pp. 940–943). IEEE.
[194] Van den Broek, E., Schut, M., Westerink, J., & Tuinenbreijer, K. (2009). Unobtrusive
Sensing of Emotions (USE). Journal of Ambient Intelligence and Smart Environments, 1(3),
287–299.
[195] Van den Broek, E., & Westerink, J. (2009). Considerations for emotion-aware consumer
products. Applied Ergonomics, 40(6), 1055–1064.
[196] Zeng, Z., Pantic, M., Roisman, G., & Huang, T. (2008). A survey of affect recognition
methods: Audio, visual, and spontaneous expressions. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 31(1), 39–58.
[197] Poels, K., & Dewitte, S. (2006). How to capture the heart? reviewing 20 years of emotion
measurement in advertising. Journal of Advertising Research-New York, 46(1), 18.
[198] Chanel, G., Kierkels, J., Soleymani, M., & Pun, T. (2009). Short-term emotion assessment
in a recall paradigm. International Journal of Human-Computer Studies, 67(8), 607–627.
[199] Healey, J. (2009). Affect detection in the real world: Recording and processing physiological
signals. In 3rd International Conference on Affective Computing and Intelligent Interaction
and Workshops, 2009 (ACII 2009) (pp. 1–6). IEEE.
[200] Lang, P., & Bradley, M. (1999). International affective digitized sounds (IADS): Stimuli, 802
instruction manual and affective ratings (Technical Report no. b-2) (Vol. 803). Gainsville.
[201] Bensafi, M. (2001). Le traitement affectif des odeurs: aspects implicites et explicites. Ph.d
Thesis in, Lyon 2 University.
[202] Lang, P., Bradley, M., & Cuthbert, B. (1997). International affective picture system
(IAPS): Technical manual and affective ratings. Gainesville: The Center for Research in
Psychophysiology, University of Florida.
[203] Lang, P., Greenwald, M., Bradley, M., & Hamm, A. (1993). Looking at pictures: Affective,
facial, visceral, and behavioral reactions. Psychophysiology, 30(3), 261–273.
[204] Lang, P. (1980). Behavioral treatment and bio-behavioral assessment: Computer applica-
tions. Technology in Mental Health Care Delivery Systems, 119–137.
[205] Grimm, S., Schmidt, C., Bermpohl, F., Heinzel, A., Dahlem, Y., Wyss, M., et al. (2006).
Segregated neural representation of distinct emotion dimensions in the prefrontal cortex–an
fMRI study. Neuroimage, 30(1), 325–340.
[206] Hariri, A., Mattay, V., Tessitore, A., Fera, F., & Weinberger, D. (2003). Neocortical
modulation of the amygdala response to fearful stimuli. Biological Psychiatry, 53(6),
494–501.
[207] Chatel-Goldman, J., Congedo, M., Jutten, C., & Schwartz, J.-L. (2014). Touch increases
autonomic coupling between romantic partners. Frontiers in Behavioral Neuroscience, 8, 95.
http://doi.org/10.3389/fnbeh.2014.00095.
Bibliography 133
[208] Liljencrantz, J., & Olausson, H. (2014). Tactile c fibers and their contributions to pleasant
sensations and to tactile allodynia. Frontiers in Behavioral Neuroscience, 8, 37. http://doi.
org/10.3389/fnbeh.2014.00037.
[209] Zotterman, Y. (1939). Touch, pain and tickling: An electro-physiological investigation on
cutaneous sensory nerves. Journal of Physiology, 95(1), 1–28.
[210] Rolls, E. T. (2010). The affective and cognitive processing of touch, oral texture, and
temperature in the brain. Neuroscience & Biobehavioral Reviews, 34(2), 237–245.
[211] Triscoli, C., Olausson, H., Sailer, U., Ignell, H., & Croy, I. (2013). CT-optimized skin
stroking delivered by hand or robot is comparable. Frontiers in Behavioral Neuroscience,
7, 208.
[212] Löken, L. S., Wessberg, J., McGlone, F., & Olausson, H. (2009). Coding of pleasant touch
by unmyelinated afferents in humans. Nature Neuroscience, 12(5), 547–548.
[213] Kessler, R., McGonagle, K., Zhao, S., Nelson, C., Hughes, M., Eshleman, S., et al. (1994).
Lifetime and 12-month prevalence of dsm-iii-r psychiatric disorders in the united states:
Results from the national comorbidity survey. Archives of General Psychiatry, 51(1), 8.
[214] Kauer-SantAnna, M., Kapczinski, F., & Vieta, E. (2009). Epidemiology and management of
anxiety in patients with bipolar disorder. CNS Drugs, 23(11), 953–964.
[215] Young, R., Biggs, J., Ziegler, V., & Meyer, D. (1978). A rating scale for mania: Reliability,
validity and sensitivity. The British Journal of Psychiatry, 133(5), 429–435.
[216] Vieta, E., Reinares, M., & Rosa, A. (2011). Staging bipolar disorder. Neurotoxicity Research,
19(2), 279–285.
[217] Andreazza, A., Kauer-Sant’Anna, M., Frey, B., Bond, D., Kapczinski, F., Young, L., et al.
(2008). Oxidative stress markers in bipolar disorder: A meta-analysis. Journal of Affective
Disorders, 111(2), 135–144.
[218] Phillips, M., & Vieta, E. (2007). Identifying functional neuroimaging biomarkers of bipolar
disorder: Toward dsm-v. Schizophrenia Bulletin, 33(4), 893–904.
[219] Cohen, H., Kaplan, Z., Kotler, M., Mittelman, I., Osher, Y., & Bersudsky, Y. (2003). Impaired
heart rate variability in euthymic bipolar patients. Bipolar Disorders, 5(2), 138–143.
[220] Henry, B. L., Minassian, A., Paulus, M. P., Geyer, M. A., & Perry, W. (2010). Heart rate
variability in bipolar mania and schizophrenia. Journal of Psychiatric Research, 44(3),
168–176.
[221] Levy, B. (2013). Autonomic nervous system arousal and cognitive functioning in bipolar
disorder. Bipolar Disorders, 15(1), 70–79.
[222] Valenza, G., Nardelli, M., Lanata, A., Gentili, C., Bertschy, G., Paradiso, R., et al., Wearable
monitoring for mood recognition in bipolar disorder based on history-dependent long-term
heart rate variability analysis. IEEE Journal of Biomedical and Health Informatics, 18(5),
1625–1635.
[223] Iverson, G., Gaetz, M., Rzempoluck, E., McLean, P., Linden, W., & Remick, R. (2005). A
new potential marker for abnormal cardiac physiology in depression. Journal of Behavioral
Medicine, 28(6), 507–511.
[224] Taillard, J., Lemoine, P., Boule, P., Drogue, M., & Mouret, J. (1993). Sleep and heart rate
circadian rhythm in depression: The necessity to separate. Chronobiology International,
10(1), 63–72.
[225] Taillard, J., Sanchez, P., Lemoine, P., & Mouret, J. (1990). Heart rate orcadian rhythm
as a biological marker of desynchronization in major depression: A methodological and
preliminary report. Chronobiology International, 7(4), 305–316.
[226] Iacono, W. G., & Tuason, V. B. (1983). Bilateral electrodermal asymmetry in euthymic
patients with unipolar and bipolar affective disorders. Biological Psychiatry, 18(3), 303–315.
[227] Iacono, W. G., Lykken, D. T., Peloquin, L. J., Lumry, A. E., Valentine, R. H., & Tuason,
V. B. (1983). Electrodermal activity in euthymic unipolar and bipolar affective disorders: A
possible marker for depression. Archives of General Psychiatry, 40(5), 557.
134 Bibliography
[228] Sponheim, S., Allen, J., & Iacono, W. (1995). Selected psychophysiological measures
in depression: The significance of electrodermal activity, electroencephalographic asym-
metries, and contingent negative variation to behavioral and neurobiological aspects
of depression. The behavioral high risk paradigm in psychopathology (pp. 222–249).
New York: Springer.
[229] Damasio, A. R. (1998). Emotion in the perspective of an integrated nervous system. Brain
Research Reviews, 26(2), 83–86.
[230] Coan, J. A., & Allen, J. J. (2007). Handbook of emotion elicitation and assessment. Oxford:
Oxford university press.
[231] Ruiz-Padial, E., Vila, J., & Thayer, J. (2011). The effect of conscious and non-conscious
presentation of biologically relevant emotion pictures on emotion modulated startle and
phasic heart rate. International Journal of Psychophysiology, 79(3), 341–346.
[232] Beissner, F., Meissner, K., Bär, K.-J., & Napadow, V. (2013). The autonomic brain: An
activation likelihood estimation meta-analysis for central processing of autonomic function.
The Journal of Neuroscience, 33(25), 10,503–10,511.
[233] Heller, A., Johnstone, T., Shackman, A., Light, S., Peterson, M., Kolden, G., et al. (2009).
Reduced capacity to sustain positive emotion in major depression reflects diminished
maintenance of fronto-striatal brain activation. Proceedings of the National Academy of
Sciences, 106(52), 22,445–22,450.
[234] Picard, R. W. (2003). Affective computing: Challenges. International Journal of Human-
Computer Studies, 59(1), 55–64.
[235] Rottenberg, J., Ray, R. R., & Gross, J. J. (in press). Emotion elicitation using films. In J. A.
Coan & J. J. B. Allen (Eds.), The handbook of emotion elicitation and assessment. New York:
Oxford University Press.
[236] Nasoz, F., Alvarez, K., Lisetti, C. L., & Finkelstein, N. (2004). Emotion recognition
from physiological signals using wireless sensors for presence technologies. Cognition,
Technology & Work, 6(1), 4–14.
[237] Pecchinenda, A. (1996). The affective significance of skin conductance activity during a
difficult problem-solving task. Cognition & Emotion, 10(5), 481–504.
[238] Scheirer, J., Fernandez, R., Klein, J., & Picard, R. W. (2002). Frustrating the user on purpose:
A step toward building an affective computer. Interacting with Computers, 14(2), 93–118.
[239] Ilves, M., & Surakka, V. (2012). Heart rate responses to synthesized affective spoken words.
Advances in Human-Computer Interaction, 2012, 14.
[240] Yang, Y.-H., & Chen, H. H. (2011). Music emotion recognition. Milton Park/Abingdon-on-
Thames/Oxfordshire: Taylor & Francis Group.
[241] Yang, Y.-H., Lin, Y.-C., Su, Y.-F., & Chen, H. H. (2008). A regression approach to music
emotion recognition. IEEE Transactions on Audio, Speech, and Language Processing, 16(2),
448–457.
[242] Kim, J., & Andre, E. (2008). Emotion recognition based on physiological changes in
music listening. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(12),
2067–2083.
[243] Ramakrishnan, S. (2012). Recognition of emotion from speech: A review. In S.
Ramakrishnan (Ed.), Speech enhancement, modeling and recognition–algorithms
and applications (p. 121). ISBN: 978-953-51-0291-5, 2012. InTech, Available from:
http://www.intechopen.com/books/speech-enhancementmodeling-and-recognition-
algorithmsandapplications/recognition-of-emotion-from-speech-areview.
[244] Levenson, R. W. (1988). Emotion and the autonomic nervous system: A prospectus for
research on autonomic specificity. Social psychophysiology and emotion: Theory and
clinical applications (pp. 17–42). Oxford: Wiley.
[245] Andreassi, J. L. (2000). Psychophysiology: Human behavior and physiological response.
New York: Psychology Press.
Bibliography 135
[246] Lanata, A., Greco, A., Valenza, G., & Scilingo, E. P. (2014). A pattern recognition approach
based on electrodermal response for pathological mood identification in bipolar disorders. In
2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
(pp. 3601–3605). IEEE.
[247] Bianchi, M., Valenza, G., Serio, A., Lanata, A., Greco, A., Nardelli, M., et al. (2014). Design
and preliminary affective characterization of a novel fabric-based tactile display. In 2014
IEEE Haptics Symposium (HAPTICS) (pp. 591–596). IEEE.
[248] Valenza, G., Lanata, A., & Scilingo, E. (2012). The role of nonlinear dynamics in affective
valence and arousal recognition. IEEE Transactions on Affective Computing, 3(2), 237–249.
[249] Bradley, M., & Lang, P. J. (1999). The International affective digitized sounds (IADS)
stimuli, instruction manual and affective ratings. NIMH Center for the Study of Emotion
and Attention. Gainesville: University of Florida.
[250] Bensafi, M., Rouby, C., Farget, V., Bertrand, B., Vigouroux, M., & Holley, A. (2002).
Autonomic nervous system responses to odours: The role of pleasantness and arousal.
Chemical Senses, 27(8), 703–709.
[251] Gibbons, J. D., & Chakraborti, S. (2011). Nonparametric statistical inference. Berlin:
Springer.
[252] Jain, A., & Zongker, D. (1997). Feature selection: Evaluation, application, and small sample
performance. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(2),
153–158.
[253] Wang, J.-C., Wang, J.-F., He, K. W., & Hsu, C.-S. (2006). Environmental sound classification
using hybrid svm/knn classifier and MPEG-7 audio low-level descriptor. In International
Joint Conference on Neural Networks, 2006 (IJCNN06) (pp. 1731–1735). IEEE.
[254] Kearns, M., & Ron, D. (1999). Algorithmic stability and sanity-check bounds for leave-one-
out cross-validation. Neural Computation, 11(6), 1427–1453.
[255] Duda, R. O., Hart, P. E., & Stork, D. G. (2012). Pattern classification. New York: Wiley.
[256] Kohavi, R., & Provost, F. (1988). Glossary of terms. Machine Learning, 30, 271–274.
[257] Lemke, M. R., Fischer, C. J., Wendorff, T., Fritzer, G., Rupp, Z., & Tetzlaff, S. (2005).
Modulation of involuntary and voluntary behavior following emotional stimuli in healthy
subjects. Progress in Neuro-Psychopharmacology and Biological Psychiatry, 29(1), 69–76.
[258] Schaaff, K., & Schultz, T. (2009). Towards emotion recognition from electroencephalo-
graphic signals. In 3rd International Conference on Affective Computing and Intelligent
Interaction and Workshops, 2009 (ACII 2009) (pp. 1–6). IEEE.
[259] Frantzidis, C. A., Bratsas, C., Papadelis, C. L., Konstantinidis, E., Pappas, C., & Bamidis,
P. D. (2010). Toward emotion aware computing: an integrated approach using multichannel
neurophysiological recordings and affective visual stimuli. IEEE Transactions on Informa-
tion Technology in Biomedicine, 14(3), 589–597.
[260] Palomba, D., Angrilli, A., & Mini, A. (1997). Visual evoked potentials, heart rate responses
and memory to emotional pictorial stimuli. International Journal of Psychophysiology,
27(1), 55–67.
[261] Junghöfer, M., Schupp, H. T., Stark, R., & Vaitl, D. (2005). Neuroimaging of emotion:
Empirical effects of proportional global signal scaling in fMRI data analysis. NeuroImage,
25(2), 520–526.
[262] Jackson Davis, W., Rahman, M. A., Smith, L. J., Burns, A., Senecal, L., McArthur, D.,
et al. (1995). Properties of human affect induced by static color slides (IAPS): Dimensional,
categorical and electromyographic analysis. Biological Psychology, 41(3), 229–253.
[263] Sloan, D. M., Bradley, M. M., Dimoulas, E., & Lang, P. J. (2002). Looking at facial
expressions: Dysphoria and facial EMG. Biological Psychology, 60(2), 79–90.
[264] Gavazzeni, J., Wiens, S., & Fischer, H. (2008). Age effects to negative arousal differ for
self-report and electrodermal activity. Psychophysiology, 45(1), 148–151.
[265] Norris, C. J., Larsen, J. T., & Cacioppo, J. T. (2007). Neuroticism is associated with
larger and more prolonged electrodermal responses to emotionally evocative pictures.
136 Bibliography
[266] Greco, A., Lanata, A., Valenza, G., Scilingo, E. P., & Citi, L. (2014). Electrodermal activity
processing: A convex optimization approach. In 2014 36th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (EMBC) (pp. 2290–2293). IEEE.
[267] Pollatos, O., Herbert, B. M., Matthias, E., & Schandry, R. (2007). Heart rate response
after emotional picture presentation is modulated by interoceptive awareness. International
Journal of Psychophysiology, 63(1), 117–124.
[268] Lang, P. (1995). The emotion probe. Studies of motivation and attention. The American
Psychologist, 50(5), 372.
[269] McCurdy, H. G. (1950). Consciousness and the galvanometer. Psychological Review, 57(6),
322.
[270] Bradley, M. M., & Lang, P. J. (2007). The international affective digitized sounds (iads):
Affective ratings of sounds and instruction manual. Technical Report B-3, University of
Florida, Gainesville.
[271] Lang, P. J., Bradley, M. M., & Cuthbert, B. N. (2005). International affective picture system
(IAPS): Affective ratings of pictures and instruction manual. Cuthbert: NIMH, Center for
the Study of Emotion & Attention.
[272] Gobl, C., & Ni, A. (2003). The role of voice quality in communicating emotion, mood and
attitude. Speech Communication, 40(1), 189–212.
[273] Goldstein, A. (1980). Thrills in response to music and other stimuli. Physiological Psychol-
ogy, 8(1), 126–129.
[274] Johna, S. (1991). Music structure and emotional response: Some empirical findings.
Psychology of Music, 991(9), L120.
[275] Marcell, M., Malatanos, M., Leahy, C., & Comeaux, C. (2007). Identifying, rating, and
remembering environmental sound events. Behavior Research Methods, 39(3), 561–569.
[276] Gaver, W. W. (1993). What in the world do we hear? An ecological approach to auditory
event perception. Ecological Psychology, 5(1), 1–29.
[277] Blood, A. J., & Zatorre, R. J. (2001) Intensely pleasurable responses to music correlate
with activity in brain regions implicated in reward and emotion. Proceedings of the National
Academy of Sciences, 98(20), 11,818–11,823.
[278] Koelsch, S., Fritz, T., Müller, K., Friederici, A. D., et al. (2006). Investigating emotion with
music: An fMRI study. Human Brain Mapping, 27(3), 239–250.
[279] Anders, S., Eippert, F., Weiskopf, N., & Veit, R. (2008). The human amygdala is sensitive to
the valence of pictures and sounds irrespective of arousal: An fMRI study. Social Cognitive
and Affective Neuroscience, 3(3), 233–243.
[280] Roque, A. L., Valenti, V. E., Guida, H. L., Campos, M. F., Knap, A., Vanderlei, L. C. M., et
al. (2013). The effects of auditory stimulation with music on heart rate variability in healthy
women. Clinics, 68(7), 960–967.
[281] Orini, M., Bailon, R., Enk, R., Koelsch, S., Mainardi, L., & Laguna, P. (2010). A method
for continuously assessing the autonomic response to music-induced emotions through HRV
analysis. Medical & Biological Engineering & Computing, 48(5), 423–433.
[282] Anttonen, J., & Surakka, V. (2005). Emotions and heart rate while sitting on a chair.
In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems
(pp. 491–499). ACM.
[283] Ivonin, L., Chang, H.-M., Chen, W., & Rauterberg, M. (2013). Unconscious emotions: Quan-
tifying and logging something we are not aware of. Personal and Ubiquitous Computing,
17(4), 663–673.
[284] Verona, E., Patrick, C. J., Curtin, J. J., Bradley, M. M., & Lang, P. J. (2004). Psychopathy and
physiological response to emotionally evocative sounds. Journal of Abnormal Psychology,
113(1), 99.
[285] Hariharan, A., & Adam, M. T. P. (2015). Blended emotion detection for decision support.
IEEE Transactions on Human-Machine Systems, 45(4), 510–517.
[286] Storm, H. (2008). Changes in skin conductance as a tool to monitor nociceptive stimulation
and pain. Current Opinion in Anesthesiology, 21(6), 796–804.
Bibliography 137
[287] Hellerud, B., & Storm, H. (2002). Skin conductance and behaviour during sensory stimula-
tion of preterm and term infants. Early Human Development, 70(1), 35–46.
[288] Olausson, H., Cole, J., Rylander, K., McGlone, F., Lamarre, Y., Wallin, B. G., et al. (2008).
Functional role of unmyelinated tactile afferents in human hairy skin: Sympathetic response
and perceptual localization. Experimental Brain Research, 184(1), 135–140.
[289] Löken, L. S., Evert, M., & Wessberg, J. (2011). Pleasantness of touch in human glabrous
and hairy skin: Order effects on affective ratings. Brain Research, 1417, 9–15.
[290] Klöcker, A., Oddo, C. M., Camboni, D., Penta, M., & Thonnard, J.-L. (2014). Physical
factors influencing pleasant touch during passive fingertip stimulation. PloS One, 9(7),
e101361.
[291] Bradley, M. M., & Lang, P. J. (1994). Measuring emotion: the self-assessment manikin and
the semantic differential. Journal of Behavior Therapy and Experimental Psychiatry, 25(1),
49–59.
[292] Posner, J., Russell, J. A., & Peterson, B. S. (2005). The circumplex model of affect: An
integrative approach to affective neuroscience, cognitive development, and psychopathology.
Development and Psychopathology, 17(03), 715–734.
[293] Herz, R. S., & Engen, T. (1996). Odor memory: review and analysis. Psychonomic Bulletin
& Review, 3(3), 300–313.
[294] Van Toller, S. (1988). Emotion and the brain. In Perfumery (pp. 121–143). Berlin: Springer.
[295] Alaoui-Ismaili, O., Robin, O., Rada, H., Dittmar, A., & Vernet-Maury, E. (1997). Basic emo-
tions evoked by odorants: Comparison between autonomic responses and self-evaluation.
Physiology & Behavior, 62(4), 713–720.
[296] Lorig, T. S., Huffman, E., DeMartino, A., & DeMarco, J. (1991). The effects of low
concentration odors on eeg activity and behavior. Journal of Psychophysiology, 5(1), 69–77.
[297] Aggleton, J. P., & Mishkin, M. (1986). The amygdala: sensory gateway to the emotions.
Emotion: Theory, Research and Experience, 3, 281–299.
[298] Price, J. L. (1987). The central olfactory and accessory olfactory systems. In T. E. Finger, &
W. L. Silver (Eds.), Neurobiology of Taste and Smell (pp. 179–203). New York: Wiley.
[299] Greco, A., Valenza, G., Nardelli, M., Bianchi, M., Citi, L., & Scilingo, E. P. (2016). Force-
velocity assessment of caress-like stimuli through the electrodermal activity processing:
Advantages of a convex optimization approach. IEEE Transactions on Human-Machine
Systems, PP(99), 1–10.
[300] Van Toller, C., Kirk-Smith, M., Wood, N., Lombard, J., & Dodd, G. (1983). Skin con-
ductance and subjective assessments associated with the odour of 5-˛-androstan-3-one.
Biological Psychology, 16(1), 85–107.
[301] Robin, O., Alaoui-Ismaili, O., Dittmar, A., & Vernet-Maury, E. (1999). Basic emotions
evoked by eugenol odor differ according to the dental experience. a neurovegetative analysis.
[302] Brauchli, P., Rüegg, P. B., Etzweiler, F., & Zeier, H. (1995). Electrocortical and autonomic
alteration by administration of a pleasant and an unpleasant odor. Chemical Senses, 20(5),
505–515.
[303] Uryvaev, Y., Golubeva, N., & Nechaev, A. (1986). Differences in human involuntary
reactions to perceptible and imperceptible odors. In Doklady Akademii Nauk SSSR, 290,
501–504.
[304] Henion, K. E. (1971). Odor pleasantness and intensity: A single dimension? Journal of
Experimental Psychology, 90(2), 275.
[305] Moskowitz, H. R., Dravnieks, A., & Gerbers, C. (1974). Odor intensity and pleasantness of
butanol. Journal of Experimental Psychology, 103(2), 216.
[306] Moskowitz, H. R., Dravnieks, A., & Klarman, L. A. (1976). Odor intensity and pleasantness
for a diverse set of odorants. Attention, Perception, & Psychophysics, 19(2), 122–128.
[307] Doty, R. L. (1975). An examination of relationships between the pleasantness, intensity, and
concentration of 10 odorous stimuli. Perception & Psychophysics, 17(5), 492–496.
138 Bibliography
[308] Alaoui-Ismaili, O., Vernet-Maury, E., Dittmar, A., Delhomme, G., & Chanel, J. (1997).
Odor hedonics: Connection with emotional response estimated by autonomic parameters.
[309] Kring, A. M., & Gordon, A. H. (1998). Sex differences in emotion: Expression, experience,
and physiology. Journal of Personality and Social Psychology, 74(3), 686.
[310] Bradley, M. M., Codispoti, M., Sabatinelli, D., & Lang, P. J. (2001). Emotion and motivation
ii: Sex differences in picture processing. Emotion, 1(3), 300.
[311] Nardelli, M., Valenza, G., Bianchi, M., Greco, A., Lanata, A., Bicchi, A., et al. (2015).
Gender-specific velocity recognition of caress-like stimuli through nonlinear analysis of
heart rate variability. In 2015 37th Annual International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC) (pp. 298–301). IEEE.
[312] Grossman, M., & Wood, W. (1993). Sex differences in intensity of emotional experience: A
social role interpretation. Journal of Personality and Social Psychology, 65(5), 1010.
[313] Robin, O., Rousmans, S., Dittmar, A., & Vernet-Maury, E. (2003). Gender influence on
emotional responses to primary tastes. Physiology & Behavior, 78(3), 385–393.
[314] Wrase, J., Klein, S., Gruesser, S. M., Hermann, D., Flor, H., Mann, K., et al. (2003).
Gender differences in the processing of standardized emotional visual stimuli in humans:
a functional magnetic resonance imaging study. Neuroscience Letters, 348(1), 41–45.
[315] Soussignan, R., & Schall, B. (1996). Children’s facial responsiveness to odors: Influences of
hedonic valence of odor, gender, age, & social presence. Developmental Psychology, 32(2),
367.
[316] Yousem, D. M., Maldjian, J. A., Siddiqi, F., Hummel, T., Alsop, D. C., Geckle, R. J., et
al. (1999). Gender effects on odor-stimulated functional magnetic resonance imaging. Brain
Research, 818(2), 480–487.
[317] Naudin, M., El-Hage, W., Gomes, M., Gaillard, P., Belzung, C., & Atanasova, B. (2012).
State and trait olfactory markers of major depression. PLoS One, 7(10), e46938.
[318] Kroenke, K., Spitzer, R., & Williams, J. (2001). The PHQ-9. Journal of General Internal
Medicine, 16(9), 606–613.
[319] Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics Bulletin, 1(6),
80–83.
[320] Siegel, S. (1956). The Mann-Whitney U test. Nonparametric statistics for the behavioral
sciences (pp. 116–127). New York: McGraw-Hill.
[321] Meli, L., Scheggi, S., Pacchierotti, C., & Prattichizzo, D. (2014). Wearable haptics and hand
tracking via an RGB-d camera for immersive tactile experiences. In ACM SIGGRAPH 2014
Posters (p. 56). ACM.
[322] Mørkrid, L., & Qiao, Z.-G. (1988). Continuous estimation of parameters in skin electrical
admittance from simultaneous measurements at two different frequencies. Medical and
Biological Engineering and Computing, 26(6), 633–640.
[323] Sawan, M., Laaziri, Y., Mounaim, F., Elzayat, E., Corcos, J., & Elhilali, M. (2007).
Electrode–tissues interface: Modeling and experimental validation. Biomedical Materials,
2(1), S7.
[324] Pavšelj, N., Préat, V., & Miklavčič, D. (2007). A numerical model of skin electroper-
meabilization based on in vivo experiments. Annals of Biomedical Engineering, 35(12),
2138–2144.
[325] Empatica: Human data in real time. https://www.empatica.com/ (2016).
[326] Fernandez, R., & Picard, R. W. (2003). Modeling drivers’ speech under stress. Speech
Communication, 40(1), 145–159.

Advances in Electrodermal Activity Processing With Applications For Mental Health

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advances in Electrodermal Activity Processing With Applications For Mental Health

Uploaded by

Copyright:

Available Formats

Alberto

Enzo Pasquale Scilingo

ISBN 978-3-319-46704-7 ISBN 978-3-319-46705-4 (eBook)

Library of Congress Control Number: 2016952862

© Springer International Publishing AG 2016

Printed on acid-free paper

This Springer imprint is published by Springer Nature

Colchester, UK Dr. Luca Citi

disorder, it is shown how EDA significantly changes according to different mood

Pisa, Italy Alberto Greco

1 Electrodermal Phenomena and Recording Techniques . . . . . . . . . . . . . . . . . . 1

2.4.3 Observation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

5.6 Affective Olfactory Elicitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

Impedance cardiogram ICG

This book is intended to provide an exhaustive description of the electrodermal

1.1 Electrodermal Activity and Skin Conductance

© Springer International Publishing AG 2016 1

Table 1.1 Structure of the Cutis Epidermis Stratum corneum

1.2 Anatomy of the Skin

1.3 Anatomy of Sweat Glands

As a matter of fact, although sweating occurs in response to extremely different

1.4 Physiology of the Electrodermal System

1.4.1 Mechanisms of the Electrodermal Electrophysiological

By applying a potential difference between two sites of the skin, it is possible

1.4.2 Genesis of the Electrodermal Response

Fig. 1.4 Sequence of events during the electrodermal response

• If the rate of secretion is high enough, it can generate a pressure intraductal

1.5 Recording Systems

Moreover, to our knowledge, no systems able to perform both DC and AC

1.5.1 Measurement Sites

As mentioned above, conventionally, EDA is measured at palmar and finger sites.

1.6 Exemplary Electrodermal Activity Monitoring Devices

So far, the hitherto nearly exclusive use of DC methods in exosomatic EDA

Fig. 1.6 Sequence of events Bipolar

1.6.1 DC Source Front-End

As mentioned, DC recording of EDA is the most widely used method of mea-

1.6.2 AC Source Front-End

1.6.3 Remote DC Devices

• Biosemi. Biosemi is a product intended to be used for research applications

1.6.4 Wearable DC Devices

Fig. 1.9 Sensorized glove

1.6.5 Multi-Frequency Sensorized Glove

As already anticipated, here we briefly preset a new wearable acquisition device,

Fig. 1.10 Block scheme of the electronic circuit

Moreover, wireless communication was implemented by an Xbee module con-

2.1 Mathematical Models of the EDA: An Overview

© Springer International Publishing AG 2016 19

In this book, we report on a recently proposed method called CvxEDA [69] to

2.2 EDA Analysis

2.2.1 Conventional Analysis

Fig. 2.1 General diagram of a conventional analysis procedure

2.2.2 Model-Based Approach

2.2.2.1 Model Evaluation

2.3 CDA: Continuous Deconvolution Analysis

The CDA is a method, proposed by Benedek et al., to decompose the skin

2.3.2 EDA Deconvolution Analysis

EDA D SMNA ˝ IRF (2.2)

where SMNA D .DRIVERtonic C DRIVERphasic /. In the Eq. (2.2), SMNA is unknown

2.4 CvxEDA: A Convex Optimization Approach

2.4.1 Convex Optimization

A set K 2 Rn is convex if:

8 x; y 2 K and 2 Œ0I 1. A function is convex if:

f .x C .1 /y/  f .x/ C .1 /f .y/ (2.4)

8 x; y 2 K and 2 Œ0I 1. A function is convex if:

f .x C .1 /y/ f .x/ C .1 /f .y/ (2.4)

Œq; `; d D arg maxq;`;d PŒq; `; d j y: (2.14)

PŒq; `; d j y / PŒy j q; `; d PŒq PŒ` PŒd; (2.15)

where PŒy j q; `; d is the likelihood of observing a specific SC time series given